Different parameters such as finance charge rate, finance term, and relative ... between finance rate and discount factor would result in different optimal ...
MICRO-GRID PORTFOLIO OPTIMIZATION UNDER UNCERTAINTY Farnaz Farzan and Mohsen A Jafari Department of Industrial & Systems Engineering Rutgers University
Abstract In this paper we propose an integrated two-step approach to micro-grid power generation portfolio optimization under uncertainty. The portfolio includes solar photovoltaic panels (PVs), wind turbine, gas-fired generation, storage and purchase from the grid. The model uniquely integrates short-term uncertainties rising from micro-grid operation, and the long-term uncertainties due to future natural gas prices, investment in renewable assets, and financing costs. This work extends the current literature in two major ways: (i) It takes a holistic approach to investment by including different types of distributed generations in micro-grid portfolio, (ii) It directly includes short-term planning and operational risks and long-term investment and pricing risks and integrates them into a single two-step optimization model. Finally, the solution approach uniquely combines a general binomial lattice with mixed integer quadratic model for budgeting and a regression model that estimates cost of operation and planning micro-grid with its current resources and load. The proposed framework allows us to study the impact of individual generation assets and their interactions on investment decisions. We are also able to quantify the impact of uncertainties and operational stochasticity on investment decisions. Key Words: Micro-grids, asset portfolio optimization, investment under uncertainty, capital budgeting
Introduction We are seeking solutions to optimal investment on a micro-grid power generation portfolio under uncertainty. Micro-grid is defined as a collection of distributed energy resources including power generation and energy storage technologies (e.g., thermal and electric battery) that can serve all or part of electric and heat demand within the same locality. More often the term micro-grid is used for such a system that has the ability to island once there is a macro-grid’s outage event. As such, the value of micro-grid is driven by savings in part of energy costs that should have been otherwise supplied from external resources such as utility companies and/or retailers. Moreover, micro-grids can lend themselves to more renewable energy and significantly reduce the need for power transmission infrastructure by producing reliable energy close to where it is needed. A micro-grid, if configured and operated properly, can lead to a cheaper, more sustainable and resilient energy supply for a community. In lieu of the recent natural catastrophic weather
1
patterns, many cities and communities in the USA, especially along the eastern border, are considering micro-grids as reliable alternatives for energy resiliency and security. In this article, a micro-grid portfolio includes solar PVs, wind turbine, gas-fired generation, storage (electric battery) and purchase from the grid. We intend to uniquely integrate into a single model the short-term uncertainties arising from micro-grid operation, and the long-term uncertainties due to future natural gas prices, investment in renewable assets, and financing costs. This work extends the current state of art in investment on distributed generation and micro-grids as follows: (i) Larger portfolio of power generation assets with options to purchase from grid or sell to the grid; (ii) Optimal selection of portfolio over the course of planning horizon; and (iii) Optimal incremental investment in each resource over the course of planning horizon. Moreover, the work extends the current literature as it solves for optimal investment decisions while considering a portfolio of electricity generation and storage assets and also captures short-term operational and long-term investment uncertainties. We are motivated by the fact that a proper mix of power generation resources and timely investment on these resources is an important design and operational planning decision for micro-grids. These decisions can significantly impact micro-grid long-term and short-term objectives, namely saving in energy costs, reliable and secure energy supply, reducing risks for grid on blackouts and brownouts, and the use of renewables in a generation portfolio. Furthermore, higher levels of exposure to the grid and market volatility can be avoided if the portfolio is optimized in response to its short-term load and market conditions. The value of the micro-grid portfolio depends on the return on investment and its growth on operational savings. For financial asset portfolios, the investment payoff depends on asset prices which are often embedded in aggregate information on operation and financial health of companies and/or industries. For the micro-grid, the investment payoff is directly linked to the operation of the physical assets, and return on investment is directly linked to how these operations are optimized in the short-term. As shown by Farzan et al. [7], the savings from a micro-grid could be significantly under- or over-estimated if the underlying risks were not taken into account. The long-term value of the micro-grid will also depend on when (in terms of market conditions) investments were made and also on the amount and investment financing costs. Different parameters such as finance charge rate, finance term, and relative relationship between finance rate and discount factor would result in different optimal investment decisions. Hence, the model proposed in this article integrates short-term and long-term risks into a single decision-making loop. The loop works as follows: (i) An optimization model of a daily microgrid operation is run to calculate a functional form of to-be-designed micro-grid, and (ii) The functional form is fed into a stochastic long-term investment model which decides when to invest on micro-grid components and expansions. The operation model is a simplified version of the model proposed by Farzan et al. [7]. The investment model is a stochastic mixed integer program (SMIP). A Monte Carlo simulation approach is taken where several sample path realizations over the course of the planning horizon are generated and a deterministic model for investment
2
optimization for each sample path is solved. At the end, probabilistic characteristics of investment decisions along with optimal cash flows are obtained over all sample paths. A micro-grid is a diverse portfolio of different energy generation resources, energy storage, and demand response and energy efficiency technologies. Cost and benefit streams and investment in such a system is tightly coupled with the operation of its resources. Literature lacks sufficient work addressing optimal investment uncertainty and optimal operation in enhanced micro-grid portfolios. Hybrid Optimization Model for Electric Renewable (HOMER) [8] is a product of the National Renewable Energy Laboratory, which evaluates design options for both off-grid and grid-connected power systems for remote, stand-alone and distributed generation applications. This tool does not include optimization, but different design configurations can be evaluated by comparing their operating cost/benefits and their investment costs. El Khattam et al. [5] studied the capacity investment in distributed generation (DG) in order to optimize the sizing and siting for DG capacity. Their objective function includes investment and operating costs as well as payment toward loss compensation. There are some elements of investment that are stochastic and ignoring this uncertainty can lead to poor results in investment decisions. Bruno et al. [3] consider the problem of optimal investment portfolio for a company that purchases, sells and distributes gas and owns a network of gas pipelines. They propose a two-stage stochastic programming model to solve the problem with stochastic demand. They also use conditional value at risk to control the variability of the decisions. Real options is another popular approach to address uncertainty and the option to delay an investment. Real options is very powerful in handling uncertainties, but its applications are limited to small-scale problems due to complexities in the solution methodology, unless numerical results are sought. Farzan and Jafari [6] present a real option model for a micro-grid with multiple sources of uncertainties. They show that the underlying partial differential equations can be solved under simple product form solutions and produce results that are close to the results from Monte Carlo simulations. Asano et al. [1] discuss investment strategies in a micro-grid consisting of a cogeneration system and renewable resources under uncertainty in natural gas price. They examine the sensitivity of optimal investment decisions to the level of uncertainty in gas prices.
Problem Statement and Preliminaries Consider a micro-grid generation resource portfolio including gas-fired generation, solar photo voltaic (PV), wind turbine (WT), electricity storage, and purchase from the grid. Our investment problem particularly aims at what capacity of each resource, if any, at each time period should be purchased within a planning horizon of (0, 𝝉). The objective is to maximize the cash flow due to investment in the micro-grid at the end of this horizon, which includes cash flows due to investments and operational savings prior to the end of horizon, and estimated projected cash flows beyond the horizon. This is a stochastic asset portfolio optimization problem under shortterm and long-term uncertainties. The investment decisions are subject to the following constraints:
3
A functional form that describes the short-term benefit growth of the micro-grid under short-term (operational) uncertainties, e.g., stochasticity of electricity demand, electricity spot price, solar intensity and wind speed Long-term uncertainty due to investment stochasticity including investment cost (e.g., PV and storage) and natural gas prices. We note that the available funds for investment dynamically changes over time. We assume that electricity price at peak is driven by natural gas prices. Constraints on micro-grid resources (i.e., on-site generation and energy storage).
•
•
•
The above problem can possibly be formulated into a single large-scale model, but such a model would be impractical for real life applications, especially when one seeks high granularity on both operation and investment decisions. Alternatively, one can decompose the two problems by first determining a functional form of micro-grid operational cost and then feeding this function into an investment optimization problem. We take the second approach with a cost function for the micro-grid defined at time t by 𝐶𝑜𝑠𝑡!",! = 𝑓(𝐼!",! , 𝐼!",! , 𝐼!",! , 𝐼!",! )
(1)
where GF, PV, WT, and ST stand for gas-fired, photovoltaic, wind turbine, and battery storage; and f( ) accounts for micro-grid uncertainty in both day-ahead planning and same day operation, and its argument vector defines the micro-grid characteristics where 𝐼!,! =
!"#"$%&' !" !"#$%&" ! !" !"#$ ! !"#$%&# !"#$%! !" !"#$ !
, 𝑖 = 𝐺𝐹, 𝑃𝑉, 𝑊𝑇, 𝑎𝑛𝑑 𝑆𝑇. The value of micro-grid is defined
on the basis of its electricity cost savings and earned revenue relative to no-micro-grid case. These terms will be explained later. We assume yearly investment decisions in terms of asset purchase. We also allow for borrowing funds for purchases and also alternative investment options for available cash. The objective is to maximize the accumulated cash at the end of the investment horizon accounting for beyond-the-horizon cash flows as well. Based on the dynamics of investment stochasticity, we run the investment model for possible random scenarios and examine the probability distribution of investment decisions. Nomenclature 𝐺𝐹𝐶𝑎𝑝
Gas-fired generation capacity (MW)
𝜖!"
Gas-fired generation unit heat rate (mmBtu/MWh)
𝑃𝑉𝐶𝑎𝑝
Average daily PV production or PV capacity (MW)
𝐶!"
PV constant
𝑆𝐼! ; ℎ = 1, 2, … , 24
Solar irradiance (𝑊/𝑚! )
𝑊𝑇𝐶𝑎𝑝
Average daily WT production or WT capacity (MW)
𝐶!"
WT constant
𝜂!"
WT efficiency
𝑊𝑆! ; ℎ = 1, 2, … , 24
Wind speed (m/s)
4
𝑆𝑇𝐶𝑎𝑝
Electricity storage charging/discharging rate (MW)
𝑆𝑇𝐷𝑢𝑟
Electricity storage duration (hrs)
𝐷! ; ℎ = 1, 2, … , 24
Hourly electricity demand (MW)
𝑔𝑟𝑖𝑑𝑐𝑜𝑚! ; ℎ = 1, 2, … , 24
Day-ahead hourly electricity commitment (MW)
𝑔𝑟𝑖𝑑𝑟𝑒𝑡𝑢𝑟𝑛!! ; ℎ
On-day hourly unused committed electricity purchase in each scenario (MW)
= 1, 2, … , 24 & 𝑠 = 1, … , 𝑁
𝑔𝑟𝑖𝑑𝑠𝑝𝑜𝑡!! ; ℎ = 1, 2, … , 24 & 𝑠 = 1, … , 𝑁
On-day hourly spot purchased electricity in each scenario (MW)
𝑠𝑏!! ; ℎ
On-day hourly electricity sold back to the grid in each scenario (MW)
= 1, 2, … , 24 & 𝑠 = 1, … , 𝑁
𝑔𝑓𝑐!! ; ℎ = 1, 2, … , 24 & 𝑠 = 1, … , 𝑁
Hourly gas-fired generation in each scenario (MW)
𝑐𝑠𝑡!! ; ℎ
Hourly storage charge in each scenario (MW)
= 1, 2, … , 24 & 𝑠 = 1, … , 𝑁
𝑑𝑠𝑡!! ; ℎ = 1, 2, … , 24 & 𝑠 = 1, … , 𝑁
Hourly storage discharge in each scenario (MW)
𝑠𝑙!! ; ℎ
Hourly storage level in each scenario (MW)
= 1, 2, … , 24 & 𝑠 = 1, … , 𝑁
𝐶!"#$#%,! ; ℎ = 1, 2, … , 24
Day-ahead hourly electricity price ($/MWh)
𝐶!"#$%&',! ; ℎ = 1, 2, … , 24
Day-ahead hourly penalty for unused committed electricity ($/MWh)
! 𝐶!"#$,! ; ℎ = 1, 2, … , 24 & 𝑠 = 1, … , 𝑁
On-day hourly electricity spot price ($/MWh)
𝐶!.!.
Natural gas price ($/mmBtu)
𝜖!"#$
Grid average heat rate for generation of electricity from gas (mmBtu/MWh)
𝛾! ; 𝑖 = 𝑠𝑝𝑟𝑖𝑛𝑔 𝑓𝑎𝑙𝑙 , 𝑠𝑢𝑚𝑚𝑒𝑟, 𝑤𝑖𝑛𝑡𝑒𝑟
Weight of each representative day for the year
𝑔𝑟𝑖𝑑𝑝𝑢𝑟𝑚𝑎𝑥
Maximum electricity purchased from the gird (MW)
𝑔𝑟𝑖𝑑𝑠𝑏𝑚𝑎𝑥
Maximum electricity sold back the gird (MW)
𝐶𝑎𝑝𝑖𝑛𝑐!,! ; 𝑖 = 𝐺𝐹, 𝑃𝑉, 𝑊𝑇, 𝑆𝑇; 𝑡 = 1, … , 𝜏
Incremental capacity purchased (MW)
𝑂𝑤𝑛𝐶𝑎𝑝𝑖𝑛𝑐! ; 𝑖 = 𝐺𝐹, 𝑃𝑉, 𝑊𝑇, 𝑆𝑇; 𝑡 = 1, … , 𝜏
Incremental PV capacity purchased and installed on own land (MW)
𝐸𝑥𝑃𝑉𝐶𝑎𝑝𝑖𝑛𝑐! ; 𝑖 = 𝐺𝐹, 𝑃𝑉, 𝑊𝑇, 𝑆𝑇; 𝑡 = 1, … , 𝜏
Incremental PV capacity purchased and installed on extra land (MW)
𝐶𝑎𝑝𝑒𝑥!",! ; 𝑖 = 𝐺𝐹, 𝑃𝑉, 𝑊𝑇, 𝑆𝑇; 𝑡 = 1, … , 𝜏
Unit Capacity Cost ($/MW)
𝐶𝐵𝑆! ; 𝑡 = 1, … , 𝜏
Cash spent to procure resource ($)
𝐵! ; 𝑡 = 1, … , 𝜏
Borrowed fund ($)
𝐶𝐼! ; 𝑡 = 1, … , 𝜏
Cash invested in alternative ($)
𝐼! ; 𝑖 = 𝐺𝐹, , 𝑊𝑇, 𝑆𝑇
Resource index (%)
𝑀𝑎𝑥𝐼! ; 𝑖 = 𝐺𝐹, , 𝑊𝑇, 𝑆𝑇
Maximum resource index (%)
𝑃𝑉𝑅𝑒𝑞𝐿𝑎𝑛𝑑
Land required for PV (acres/MW)
𝐿𝑎𝑛𝑑𝑃𝑟
Land price ($/acres)
𝐿𝑎𝑛𝑑𝐷𝑒𝑛𝑠
Portion of land covered by PV (%)
𝐴𝑣𝐿𝑎𝑛𝑑
Own available land for PV (acres)
𝑁𝑃𝑉
Net present value ($)
𝐶𝐹! ; 𝑡 = 1, … , 𝜏
Cash flow ($)
𝐶𝐵𝑆! ; 𝑡 = 1, … , 𝜏
Cash spent to procure resource ($)
𝐵! ; 𝑡 = 1, … , 𝜏
Borrowed fund ($)
𝐶𝐼! ; 𝑡 = 1, … , 𝜏
Cash invested in alternative ($)
5
𝐶𝐵! ; 𝑡 = 1, … , 𝜏
Cash balance ($)
𝑀𝐺𝑆𝑎𝑣𝑖𝑛𝑔! ; 𝑡 = 1, … , 𝜏
Micro-grid savings ($)
𝑅𝑂𝐼
Investment rate of return in an alternative investment (%)
𝐹𝐶
Finance charge (%)
𝐹𝑇
Finance term (years)
𝐵𝐿𝑖𝑚𝑖𝑡
Maximum Borrowing Limit ($)
Solution Approach Two models will be formulated and solved together, namely, (i) Micro-grid cost model based on Eq. (1) and (ii) Capital budgeting model. We first identify appropriate cost terms that must be included in the micro-grid cost model (see Eq. (11)). We then estimate these cost terms using a regression model (see Eq. (12)), which is defined in terms of several parameters (see Eqs. (2)(7)) that characterize a micro-grid. Regression parameters are then estimated from an optimal micro-grid operation model that takes into account short-term uncertainties, such price of fuel and electricity, and load. The regression model is then fed into a capital budgeting model. Micro-‐grid Characterization Equation (1) is defined on the basis of asset characteristics of micro-grid. Each asset is characterized by one or more parameters: Gas-fired generator is specified with its capacity of electricity generation and its heat rate. We assume a fixed heat rate and an index 𝐼!" , which represents its unit capacity: 𝐼!" =
!"#
%$(2)
![!]
PV electricity production at each hour 𝑔𝑝𝑣! is assumed to be: 𝑔𝑝𝑣! = 𝐶!" ×𝑆𝐼! (3) where 𝐶!" is PV constant and 𝑆𝐼! is solar intensity at each hour. We define a PV index, which corresponds to PV capacity: 𝐼!" =
!"#$%&# !"#$% !" !"#$%&'$'%( !"#$%&'(#) !"#$%&# !"#$% !"#$%&
=
!"#$% ![!]
=
!!! ×![!"] ![!]
(4)
where 𝐸[𝐷] and 𝐸[𝑆𝐼] are daily expected values for electricity demand and solar intensity. For wind turbine electricity generation, we use expression ([13]): 𝑔𝑤𝑡! = 𝐶!" ×𝜂!" ×𝑊𝑆!! (5) where 𝐶!" and 𝜂!" are wind turbine constant and efficiency respectively. A representative index for WT capacity is defined by: 𝐼!" =
!"#$%&# !"#$% !" !"#$%&'$'%( !"#$%&'(#) !"#$%&# !"#$% !"#$%&
=
!"#$% ![!]
=
!!" ×!!" ×![!" ! ] ![!]
(6)
To avoid higher orders in investment optimization, we assume fixed efficiency for wind turbine. We should note that in theory, wind production is proportional to wind speed cubic, however, 6
such a relationship is highly theoretical and there are no generators that can operate over a range of velocities and harvest the wind energy in that proportion (see [13]). Finally, electricity storage is characterized by two parameters, namely, its charging/discharging rate 𝑆𝑇𝐶𝑎𝑝 (MW) (assumed to be the same for charge and discharge) and its charging duration 𝑆𝑇𝐷𝑢𝑟 (hr). We assume that storage assets all have the same charging duration. The storage representative index is therefore: 𝐼!" =
!"#$% ![!]
(7)
Note that investment decision must be made per each period between t = 1,..., τ . Therefore, capacity variables, namely, 𝐺𝐹𝐶𝑎𝑝! , 𝑃𝑉𝐶𝑎𝑝! , 𝑊𝑇𝐶𝑎𝑝! , and 𝑆𝑇𝐶𝑎𝑝! are all function of t. Calculation of 𝑪𝒐𝒔𝒕𝑴𝑮,𝒕 Here we will derive Eq. (1) assuming a micro-grid with renewable assets, storage and with access to an external power grid. Moreover, the micro-grid can sell back to the grid if it is economically profitable. It is assumed that micro-grid is subject to several sources of variations: (i) variations in weather forecast, which leads to variation in the availability of renewable resources, (ii) variations in demand, and (iii) variation in spot prices. Peak electricity price on each day is assumed to be driven by natural gas price, i.e., 𝑃𝑒𝑎𝑘 𝐸𝑙𝑒𝑐 𝑃𝑟𝑖𝑐𝑒 = 𝑅𝑎𝑡𝑖𝑜!" ×𝐶!.!. ×𝜖!"#$
(8)
where 𝑅𝑎𝑡𝑖𝑜!" > 1 accounts for transmission and distribution cost at grid level and 𝜖!"#$ is grid average heat rate for generation of electricity from natural gas. Assuming a daily profile for dayahead electricity price (𝑃𝑟𝑜𝑓𝑖𝑙𝑒!"#$%&,! ) as a percentage of peak price, hourly electricity price over the course of a day is obtained by: 𝐶!"#$#%,! = 𝑃𝑒𝑎𝑘 𝐸𝑙𝑒𝑐 𝑃𝑟𝑖𝑐𝑒×𝑃𝑟𝑜𝑓𝑖𝑙𝑒!"#$%&,! (9) Next-day spot prices, electricity demand, solar radiation and wind speed are assumed to have distributions with mean and variance estimated from historical data. These random variables are correlated in their mean values but not in their variances. End-user daily electricity demand, wind speed and solar intensity profiles are inputs to the model. The annual net cost of micro-grid operation deducts any revenue from the electricity sell-back to the macro grid, computed at spot prices. Moreover, there are no operation cost of PV, WT and storage. Any planned purchase made by the micro-grid is calculated at a day-ahead price, whereas spot purchases are charged at spot prices. We allow for later modification of purchase commitment by paying a penalty as a percentage of pre-set prices. We formulate the micro-grid planning and operation optimization problem as a Two-stage stochastic programming problem [11]. A synopsis of the decisions made in the two stages follows: •
In the first stage, day-ahead plans are made to commit to the grid for a certain amount of purchase. The decision is made taking into account all sources of uncertainty. 7
•
The second stage includes observation of realized operational scenarios and taking recourse decisions for each scenario. The recourse decisions are made in terms of: a. How much of the prior commitment should really be purchased (𝑔𝑟𝑖𝑑𝑟𝑒𝑡𝑢𝑟𝑛!! ); b. How much spot electricity should be purchased (𝑔𝑟𝑖𝑑𝑠𝑝𝑜𝑡!! ) c. How much electricity from gas-fired unit should be generated (𝑔𝑓𝑐!! ). d. How much electricity should be charged to storage (𝑐𝑠𝑡!! ). e. How much electricity should be dis-charged from storage (𝑑𝑠𝑡!! ). f. How much electricity should be sold to the grid (𝑠𝑏!! ).
These recourse decisions are corrective actions to the first stage decisions for each hour depending on which random scenario is realized. We use risk-neutral two-stage stochastic programming framework to solve the micro-grid planning/operation. The model presented here takes advantage of the stochastic programming approach presented in Farzan et al. [7] and extends their results to include storage and sell back to the grid. Objective Function The objective function of this regime is the sum of the cost of first stage decision and expected net cost (i.e., cost minus revenue) of the second stage. We have: !" !!! 𝐶!"#$#%,! ×𝑔𝑟𝑖𝑑𝑐𝑜𝑚!
+
! !" ! !!![𝐶!"#$ ! ×𝑔𝑟𝑖𝑑𝑠𝑝𝑜𝑡! + 𝐶!.!. ×𝜖!" ×𝑔𝑓𝑐!! − 𝐶!"#$ !! ×𝑠𝑏!! ])
! !!! 𝑃! ×(
𝑔𝑟𝑖𝑑𝑟𝑒𝑡𝑢𝑟𝑛!! +
(𝐶!"#$%&',! −𝐶!"#$#%,! )× (10)
where the first term in Eq. (10) is the first stage cost of committing to the grid, and the other terms are the second stage costs, and P! is the probability of each scenario which is assumed to be 1/N over a discrete sampling path. Constraints The prior commitment at each time should be more than a certain value and cannot exceed a certain limit 𝑔𝑟𝑖𝑑𝑝𝑢𝑟𝑚𝑎𝑥: 𝑔𝑟𝑖𝑑𝑚𝑖𝑛 ≤ 𝑔𝑟𝑖𝑑𝑐𝑜𝑚! ≤ 𝑔𝑟𝑖𝑑𝑝𝑢𝑟𝑚𝑎𝑥 ℎ = 1, … ,24 (C1) Scenario-based constraints for a specific scenario are relevant only for the recourse decisions in that scenario and other scenario-based constraints become irrelevant for these decisions. For example, the equivalent set of constraints for spot purchase (which is a recourse decision and scenario-dependent) and sell back are dependent on scenario s: 𝑔𝑟𝑖𝑑𝑚𝑖𝑛 ≤ 𝑔𝑟𝑖𝑑𝑠𝑝𝑜𝑡!! ≤ 𝑔𝑔𝑟𝑖𝑑𝑚𝑎𝑥 ℎ = 1, … ,24 𝑎𝑛𝑑 𝑠 = 1, … , 𝑁 (C2) 𝑠𝑏!! ≤ 𝑔𝑟𝑖𝑑𝑠𝑏𝑚𝑎𝑥
(C3)
Similarly, the amount of electricity not purchased from the grid cannot exceed the prior commitment: 𝑔𝑟𝑖𝑑𝑟𝑒𝑡𝑢𝑟𝑛!! ≤ 𝑔𝑟𝑖𝑑𝑐𝑜𝑚! ℎ = 1, … ,24 𝑎𝑛𝑑 𝑠 = 1, … , 𝑁 (C4)
8
and the total purchase from the grid at each hour should not exceed the grid maximum purchase limit: 𝑔𝑟𝑖𝑑𝑐𝑜𝑚! − 𝑔𝑟𝑖𝑑𝑟𝑒𝑡𝑢𝑟𝑛!! + 𝑔𝑟𝑖𝑑𝑠𝑝𝑜𝑡!! ≤ 𝑔𝑟𝑖𝑑𝑝𝑢𝑟𝑚𝑎𝑥 ℎ = 1, … ,24 𝑎𝑛𝑑 𝑠 = 1, … , 𝑁 (C5) The electricity generation from the GF is constrained to the minimum operation level and maximum capacity of GF. Since there is no cost associated to the operation of PV and WT, electricity production from these resources is dictated by the availability of renewable resources. See Frazan et al. [7] for more details and constraints on GF, WT, PV and overall energy balance for each scenario. At the end of each time period (i.e., hour), the available energy kept in storage is conserved by: ! 𝑠𝑙!! = 𝑠𝑙!!! + 𝑐𝑠𝑡!! − 𝑑𝑠𝑡!! ℎ = 1, … ,24 𝑎𝑛𝑑 𝑠 = 1, … , 𝑁 (C6) ! where 𝑐𝑠𝑡! and 𝑑𝑠𝑡!! are charging and discharging quantities from storage during each hour. Storage charge and discharge are constrained by maximum charge/discharge rate of the device, namely,
𝑐𝑠𝑡!! ≤ 𝑆𝑇𝐶𝑎𝑝 ℎ = 1, … ,24 𝑎𝑛𝑑 𝑠 = 1, … , 𝑁 (C7) 𝑑𝑠𝑡!! ≤ 𝑆𝑇𝐶𝑎𝑝 ℎ = 1, … ,24 𝑎𝑛𝑑 𝑠 = 1, , 𝑁 (C8) Moreover, charge quantity is constrained by the remaining space left in the storage and discharge quantity is constrained by the available energy in the storage from the previous hour: ! 𝑐𝑠𝑡!! ≤ (𝑆𝑇𝐶𝑎𝑝×𝑆𝑇𝐷𝑢𝑟 − 𝑠𝑙!!! )/𝑆𝑇𝐷𝑢𝑟 ℎ = 1, … ,24 𝑎𝑛𝑑 𝑠 = 1, … , 𝑁 (C9) ! 𝑑𝑠𝑡!! ≤ 𝑠𝑙!!! /𝑆𝑇𝐷𝑢𝑟 ℎ = 1, … ,24 𝑎𝑛𝑑 𝑠 = 1, … , 𝑁 (C10)
Finally, the amount of energy stored in the device cannot exceed the maximum energy limit: 𝑠𝑙!! ≤ 𝑆𝑇𝐶𝑎𝑝×𝑆𝑇𝐷𝑢𝑟 ℎ = 1, … ,24 𝑎𝑛𝑑 𝑠 = 1, … , 𝑁 (C11) The last set of constraint indicates that selling back should be supplied from either on-site generation or discharge from storage at each hour: 𝑠𝑏!! ≤ 𝑔𝑓𝑐!! + 𝑔𝑤𝑡!! + 𝑔𝑝𝑣!! + 𝑑𝑠𝑡!! ℎ = 1, … ,24 𝑎𝑛𝑑 𝑠 = 1, … , 𝑁 (C12) The above formulation can be extended to solve for more than one day. However, to avoid lengthy computations and to demonstrate the concept, we use a representative model on the basis of three representative days over a year and extrapolate the annual cost, 𝐶𝑜𝑠𝑡!",! , based on their respective costs and weight factors: 𝐶𝑜𝑠𝑡!",! = 𝛾!"#$%&/!"## ×𝐶𝑜𝑠𝑡!",!"#$%&/!"## + 𝛾!"##$% ×𝐶𝑜𝑠𝑡!",!"##$% + 𝛾!"#$%& ×𝐶𝑜𝑠𝑡!",!"#$%&
(11)
Solving the above optimization problem for combinations of inputs defined on the basis of a proper design of experiment yields the following functional form:
9
𝐶𝑜𝑠𝑡!",! = 𝛽!,! + 𝛽!,! 𝐼!",! + 𝛽!,! 𝐼!",! + 𝛽!,! 𝐼!",! + 𝛽!,! 𝐼!",! + 𝛽!,! 𝐼!",! ×𝐼!",! + 𝛽!,! 𝐼!",! × 𝐼!",! + 𝛽!,! 𝐼!",! ×𝐼!",! + 𝛽!,! 𝐼!",! ×𝐼!",! + 𝛽!,! 𝐼!",! ×𝐼!",! + 𝛽!",! 𝐼!",! ×𝐼!",! In the above equation 𝛽!,! can be interpreted as the cost of electricity supplied by the grid. The other terms refer to micro-grid’s cost saving or revenue resulting from on-site resources: 𝑀𝐺𝑆𝑎𝑣𝑖𝑛𝑔! = 𝛽!,! 𝐼!",! + 𝛽!,! 𝐼!",! + 𝛽!,! 𝐼!",! + 𝛽!,! 𝐼!",! + 𝛽!,! 𝐼!",! ×𝐼!",! + 𝛽!,! 𝐼!",! ×𝐼!",! + 𝛽!,! 𝐼!",! ×𝐼!",! + 𝛽!,! 𝐼!",! ×𝐼!",! + 𝛽!,! 𝐼!",! ×𝐼!",! + 𝛽!",! 𝐼!",! ×𝐼!",! (12)
Capital Budgeting Model We include borrowing and lending opportunities and this leads to a more realistic operational model (Park et al. [10]). A variety of criterion functions can be optimized, including horizon models (i.e., Weingartner [12]) and internal rate of return based on selected projects. Without dwelling too much on the details of any specific modeling approach, we select a general horizon model and maximize the cash flow at the end of horizon plus the value of beyond-horizon cash flows. The model includes incremental investment decisions to form a micro-grid over a specific horizon. The decisions are made on what capacity of each resource (i.e., GF, PV, WT and electricity storage), if any, should be purchased at each time period (i.e., one year). Within each period, the active micro-grid yields a payoff in the form of cost savings and possible revenue. There would be initial cash available in the first period and it is assumed that in the beginning of each period any available cash can be either used to purchase assets or to spend in other investment opportunities. We also assume that cash inflow resulting from micro-grid’s revenue will be added to the available cash in each period. The following assumptions are made: •
•
•
•
We consider only one decision variable for each on-site generation. This is to avoid high orders in the optimization model. For example, WT efficiency and electricity storage duration are considered fixed. An asset purchased in a period will only be active at the beginning of the next period. The life of assets is assumed to be infinite and no asset depreciation is considered. Borrowing is available at a constant finance charge for a constant financing period. Funds borrowed in each period can only be invested in the same period. Outside investments are available at a fixed rate of return, and any cash invested outside in a period can be reinvested in the next period. We assume a fixed maximum borrowing limit in each period. There is a maximum limit on installed capacity of GF, WT and electricity storage. Certain land space is available for PV installation; additional capacity could be installed by paying for extra space needed. Investment rate of return in lending or any project other than microgrid assets will be less than the finance charges.
The formal presentation of the model follows:
10
Investment Optimization Model We now formulate a mixed integer quadratic programming model of the investment decision. Input variables are categorized in three groups: “On-site Resources”, “Financial Parameters” and “Resource Investment Limits”. Tables 1- 3 list input variables:
Table 1 -‐ On-‐Site Resources WT efficiency
𝜂!"
Electricity Storage Duration
𝑆𝑇𝐷𝑢𝑟
Table 2 -‐ Financial& Cost Parameters Initial Cash Available Investment Rate of Return (in an alternative investment)
Table 3 -‐ Resource Investment Limits
𝐶𝐵! ($) 𝑅𝑂𝐼 (%)
Finance Charge
𝐹𝐶 (%)
Finance Term
𝐹𝑇 (𝑦𝑒𝑎𝑟𝑠)
Maximum Borrowing Limit
𝐵𝐿𝑖𝑚𝑖𝑡($)
Unit Capacity Cost of GF
𝐶𝑎𝑝𝑒𝑥!",! ($/𝑀𝑊)
Unit Capacity Cost of PV
𝐶𝑎𝑝𝑒𝑥!",! ($/𝑀𝑊)
Unit Capacity Cost of WT
𝐶𝑎𝑝𝑒𝑥!",! ($/𝑀𝑊)
Unit Capacity Cost of ST
𝐶𝑎𝑝𝑒𝑥!",! ($/𝑀𝑊)
Price of Extra Land for PV
𝐿𝑎𝑛𝑑𝑃𝑟 ($/𝑎𝑐𝑟𝑒𝑠)
Portion of Land Covered by PV
Maximum GF Index
𝑀𝑎𝑥𝐼!"
Maximum WT Index
𝑀𝑎𝑥𝐼!"
Maximum Storage Index
𝑀𝑎𝑥𝐼!"
Owned Available Land PV Land Required for PV
𝐴𝑣𝐿𝑎𝑛𝑑 (𝑎𝑐𝑟𝑒𝑠) 𝑃𝑉𝑅𝑒𝑞𝐿𝑎𝑛𝑑 (𝑎𝑐𝑟𝑒𝑠/𝑀𝑊)
𝐿𝑎𝑛𝑑𝐷𝑒𝑛𝑠(%)
The following are the decision variables in the model: Table 4 – Decision Variables Incremental Capacity Purchased (MW)
𝐶𝑎𝑝𝑖𝑛𝑐!,! ; 𝑖 = 𝐺𝐹, 𝑃𝑉, 𝑊𝑇, 𝑆𝑇; 𝑡 = 1, … , 𝜏
Incremental Capacity purchased and installed on owned land (MW)
𝑂𝑤𝑛𝐶𝑎𝑝𝑖𝑛𝑐!,! ; 𝑖 = 𝐺𝐹, 𝑃𝑉, 𝑊𝑇, 𝑆𝑇; 𝑡 = 1, … , 𝜏
Incremental Capacity purchased and installed on extra land (MW)
𝐸𝑥𝑃𝑉𝐶𝑎𝑝𝑖𝑛𝑐!,! ; 𝑖 = 𝐺𝐹, 𝑃𝑉, 𝑊𝑇, 𝑆𝑇; 𝑡 = 1, … , 𝜏
Funds spent to procure resources in period t ($)
𝐶𝐵𝑆! ; 𝑡 = 1, … , 𝜏
11
Borrowed Funds in period t ($)
𝐵! ; 𝑡 = 1, … , 𝜏
Cash invested in other alternatives in period t ($)
𝐶𝐼! ; 𝑡 = 1, … , 𝜏
Objective function For micro-grid’s investment, the objective is to maximize the end of horizon cash flow plus the horizon time value of any cash flows beyond the horizon. End-of-horizon cash flow is the net of cash inflow and outflow at 𝜏: 𝐶𝐹! = 𝐶𝐼𝐹! − 𝐶𝑂𝐹! = 𝑀𝐺𝑆𝑎𝑣𝑖𝑛𝑔! + 𝑅𝑂𝐶! + 𝐵! − 𝐹𝑃! − 𝐶𝐵𝑆! − 𝐶𝐼! (13)
where: 𝐶𝐹! is the net cash flow at the end of the horizon 𝑀𝐺𝑆𝑎𝑣𝑖𝑛𝑔! is the cumulative microgrid savings at the end of the horizon (cash inflow) 𝑅𝑂𝐶! is the invested cash position plus return on cash invested in an alternative investment other than micro-grid in the previous period (cash inflow) 𝐵! is the borrowed fund in the last period (cash inflow) 𝐹𝑃! is the total financial charges on borrowed funds at the horizon (cash outflow) 𝐶𝐵𝑆! is the funds spent to purchase resources at the horizon (cash outflow) 𝐶𝐼! is the cash invested in alternatives projects (cash outflow). Our numerical studies show that the second order terms in Eq. (12) sufficiently explain the interactions of resources. Therefore, we will ignore higher orders. 𝑅𝑂𝐶! is the return on cash invested in an alternative investment other than micro-grid in the previous period: 𝑅𝑂𝐶! = 𝐶𝐼!!! × 1 + 𝑅𝑂𝐼
(14)
Fund borrowed in period 𝑗 would entitle the borrower to payment flows in coming periods. This payment is calculated according to normal annuity: !"×!
𝑃!,! = !!(!!!")!!!" 𝑡 = 𝑗 + 1, … , 𝑗 + 𝐹𝑇
(15)
Therefore, the total payment in each period would be: 12
𝐹𝑃! =
!!! !!! 𝑃!,! 𝑡
= 𝑗 + 1, … , 𝑗 + 𝐹𝑇
(16)
Beyond-horizon cash flows are discounted to obtain their horizon time value, 𝐶𝐹! , and include perpetual savings from the micro-grid: !
𝑀𝐺𝑆𝑎𝑣𝚤𝑛𝑔! = !!!"
! !"#$%&'(!!! !!! (!!!")!
=
!"#$%&'(!!!
(17)
!"
The return on cash invested in the last period: !
𝑅𝑂𝐶!!! = (!!!") 𝑅𝑂𝐶!!! (18) And finally the remainder of finance charges: 𝐹𝑃! =
!"! !!!" !!!!! (!!!")!
(19)
The objective is therefore given by: max (𝐶𝐹! + 𝐶𝐹! ) = max {𝑀𝐺𝑆𝑎𝑣𝑖𝑛𝑔! + 𝑅𝑂𝐶! + 𝐵! − 𝐹𝑃! − 𝐶𝐵𝑆! − 𝐶𝐼𝜏 +𝑀𝐺𝑆𝑎𝑣𝚤𝑛𝑔! + 𝑅𝑂𝐶!!! − 𝐹𝑃! } (20) Constraints - Cash flow in each period is calculated by: 𝐶𝐹! = 𝐶𝐼𝐹! − 𝐶𝑂𝐹! = 𝑀𝐺𝑆𝑎𝑣𝑖𝑛𝑔! + 𝑅𝑂𝐶! + 𝐵! − 𝐹𝑃! − 𝐶𝐵𝑆! − 𝐶𝐼𝑡 𝑡 = 1, … , 𝜏 (C13) 𝑅𝑂𝐶! is the return in period 𝑡 on cash invested outside in the previous period and is given by Eq. (16). There is a fixed maximum limit on the amount to be borrowed in each period: 𝐵! ≤ 𝐵𝑙𝑖𝑚𝑖𝑡 (C14) There is a constraint for cash invested outside based on the availability of cash: 𝐶𝐼! ≤ 𝐶𝐵! − 𝐶𝐵𝑆! (C15) 𝐶𝐼! ≤ 𝑅𝑂𝐶! + 𝑀𝐺𝑅𝑒𝑣𝑒𝑛𝑢𝑒! − 𝐶𝐵𝑆! 𝑡 = 2, … , 𝜏
(C16)
𝑀𝐺𝑅𝑒𝑣𝑒𝑛𝑢𝑒! is the micro-grid revenue from selling back to the grid. In other words, if microgrid’s cost is negative, then it is actually making money. We define a binary variable 𝐵𝑅! to determine whether micro-grid’s cost is negative in each period: 𝐵𝑅! =
1 𝐶𝑜𝑠𝑡!",! < 0 𝑡 = 1, … , 𝜏 0 𝐶𝑜𝑠𝑡!",! ≥ 0
Funds borrowed are restricted to be used to purchase on-site resources and cannot be invested on an alternative option. Therefore, total investment in micro-grid in each period is constrained by funds available through borrowing and amount spent from available cash: 𝑃𝑉𝑇𝑜𝑡𝐶𝑎𝑝𝑒𝑥! +
! !!! 𝐶𝑎𝑝𝑒𝑥!,!
×𝐶𝑎𝑝𝑖𝑛𝑐!,! = 𝐵! + 𝐶𝐵𝑆! 𝑡 = 1, … , 𝜏 𝑎𝑛𝑑 𝑖 = 𝐺𝐹, 𝑊𝑇, 𝑆𝑇 (C17)
where 𝑃𝑉𝑇𝑜𝑡𝐶𝑎𝑝𝑒𝑥! is total photovoltaic investment cost at each period. The reason for treating PV separate from other resources is due to the fact that available land for PV installation at no
13
cost imposes another constraint on PV. We assume that micro-grid owns some land for which there is no alternative value. If land requirement for the PV capacity to be installed exceeds initial available no-cost land, extra space should be acquired at a specific price to install PV excessive capacity. To be able to formulate this constraint linearly, we decompose incremental investment in PV capacity into two separate decisions, namely, incremental capacity on own land (𝑂𝑤𝑛𝐶𝑎𝑝𝑖𝑛𝑐! ) and incremental capacity in extra land (𝐸𝑥𝑃𝑉𝐶𝑎𝑝𝑖𝑛𝑐! ): 𝐶𝑎𝑝𝑖𝑛𝑐!,! = 𝑂𝑤𝑛𝐶𝑎𝑝𝑖𝑛𝑐! + 𝐸𝑥𝑃𝑉𝐶𝑎𝑝𝑖𝑛𝑐! 𝑡 = 1, … , 𝜏 (C18) As mentioned earlier, it is assumed that resources purchased in each period will not be active until the next period, therefore, in each period, the operating capacity of each resource is: 𝐶𝑎𝑝!,! = 𝐶𝑎𝑝!,!, 𝑖 = 𝐺𝐹, 𝑃𝑉, 𝑊𝑇, 𝑆𝑇 (C19) 𝐶𝑎𝑝!,! = 𝐶𝑎𝑝!,!!! + 𝐶𝑎𝑝𝑖𝑛𝑐!,!!! , 𝑡 = 2, … , 𝜏 + 1 𝑎𝑛𝑑 𝑖 = 𝐺𝐹, 𝑃𝑉, 𝑊𝑇, 𝑆𝑇 (C20) There is also a limit on the installed capacity in each period for GF, WT and storage: 𝐶𝑎𝑝!,! ≤ 𝑀𝑎𝑥𝐼! 𝑡 = 1, … , 𝜏 𝑎𝑛𝑑𝑖 = 𝐺𝐹, 𝑊𝑇, 𝑆𝑇 (C21) To keep track of how much excessive PV capacity we are allowed to invest before exceeding our own land availability, we need the following constraint: 𝐶𝑎𝑝!,!!! + 𝑂𝑤𝑛𝐶𝑎𝑝𝑖𝑛𝑐! ≤ (𝐴𝑣𝐿𝑎𝑛𝑑 ×𝐿𝑎𝑛𝑑𝐷𝑒𝑛𝑠)/𝑃𝑉𝑅𝑒𝑞𝐿𝑎𝑛𝑑 𝑡 = 1, … , 𝜏 (C22) The investor should pay an additional cost on top of investment cost for PV. Therefore the total capital cost for PV should include the price of extra land: 𝑃𝑉𝑅𝑒𝑞𝐿𝑎𝑛𝑑 ×𝐿𝑎𝑛𝑑𝑃𝑟, 𝐿𝑎𝑛𝑑𝐷𝑒𝑛𝑠 𝑡 = 1, … , 𝜏 (𝐶23)
𝑃𝑉𝑇𝑜𝑡𝐶𝑎𝑝𝑒𝑥! = 𝐶𝑎𝑝𝑒𝑥!! ×𝐶𝑎𝑝𝑖𝑛𝑐!,! + 𝐸𝑥𝑃𝑉𝐶𝑎𝑝𝑖𝑛𝑐! ×
There are additional non-negativity constraints for variables that cannot take negative values: 𝐶𝑎𝑝𝑖𝑛𝑐!,! ≥ 0; 𝑂𝑤𝑛𝐶𝑎𝑝𝑖𝑛𝑐! ≥ 0 ; 𝐶𝐵𝑆! ≥ 0; 𝐵! ≥ 0 ; 𝐶𝐼! ≥ 0; 𝑡 = 1, … , 𝜏 (C24) And constraints for binary variables: 𝐵𝑅! ∈ {0,1} 𝑡 = 1, … , 𝜏 (C25) Solution approach using Stochastic Scenario Generation To account for uncertainty, the above investment model is solved under different stochastic scenarios. Natural gas price, PV and storage costs are considered to be random variables. Sample path or realizations rising from the underlying processes are then constructed over the investment horizon divided into years. Next we describe the uncertainty dynamics of investment variables along with the scenario generation.
14
Dynamics of Natural Gas Price: A Symmetric Lattice Binomial Approach Volatile gas prices impact the investment timing and GF capacity. In this work, we assume a simple geometric Brownian motion for gas price: 𝑑𝐶 = 𝛼!.!. 𝐶𝑑𝑡 + 𝜎!.!. 𝐶𝑑𝑍!.!. (21) where 𝐶 is natural gas price ($/mmBtu), α!.!. is the natural gas annual percentage growth rate and σ!.!. is the natural gas annual percentage volatility. 𝑍! is standard Brownian motion (GBM) and 𝑑𝑍!!.!. = 𝑒 𝑑𝑡 and 𝑒~𝑁(0,1). A popular approach used to model one-factor Markov processes is the Lattice of Cox, Ross & Rubinsterin [4]. Their methodology builds a symmetrical lattice using both the deterministic and variance part of Eq. (21). Their approach therefore converges weakly to a GBM process. The expected value expression can be written as (“N. G.” is dropped for simplicity): 𝐸 C! = C! 𝑒 !!
!!!!
(22)
If we take 𝑥! = 𝐿𝑛 C! and ∆𝑡 = 𝑡 − 𝑡! , we have 𝐸[𝑥! ] = 𝑥!!! + (𝛼! −
!!! !
)∆𝑡 and 𝑉𝑎𝑟(𝑥! ) = 𝜎!! ∆𝑡
(23)
A binomial step of ∆𝑡 is considered where 𝑢 and 𝑑 are multipliers associated to up and down movements of price in each step with probabilities of 𝑝 and 1 − 𝑝 as shown in Figure 1. The following values are proposed by Cox et al. to match the first and second moments of GBM model:
Figure 1: GBM binomial step
𝑢 = 𝑒 !!
∆!
; 𝑑 = 𝑒 !!!
∆!
!
!
= ! ; 𝑝 = ! (1 +
(!! !!!! /!) !!
∆𝑡 =
!!!! !! !!!
(24)
The expected value then becomes: 𝐸 C! = 𝐶! ×𝑢×𝑝 + 𝐶! ×𝑑×(1 − 𝑝) (25) The drift parameter of GBM is presented in up and down probabilities and the lattice values model the volatility of the process. For a risk neutral approach, these probabilities are adjusted.
15
In this work we use an alternative approach proposed by Bastian-Pinto et al [2], which is equivalent to Cox et al., but applies to more general stochastic processes. Furthermore, their model is more intuitive compared to a similar model proposed by Nelson and Ramswamy [9]. The main principle is the same, i.e., one matches the first and second moments of the underlying process with the lattice parameters as follows: the deterministic expression (the first moment) of the process gives the lattice mean value and the volatility (second moment) defines the lattice up and down movements. Figure 2 illustrates the idea, where we assume that 𝑥! = 𝑥!∗ + 𝑥!! , where ! 𝑥!! is the expected value given by 𝑥!! = 𝑥!!! + (𝛼! −
!!! !
)∆𝑡 and 𝑥!∗ is the value of additive lattice
which models an arithmetic Brownian motion with zero drift and with 𝑢 and 𝑑 as its up and down increments.
Figure 2: Symmetrical binomial lattice
Dynamics of PV and Storage Investment Cost: Discrete Probability Distributions There is no specific stochastic process for PV and electricity storage investment cost. Since there is no sufficient historical data to estimate a stochastic process for PV and electricity storage investment cost over time, we assume a decreasing trend and assign a binomial probability mass function to the rate by which the investment cost decreases by each year: 𝐼!" 𝑡 =∝!" 𝐼!" 𝑡 − 1 ∝!" =
(26)
𝐷!",! 𝑤𝑖𝑡ℎ 𝑃𝑟𝑜𝑏𝑎𝑏𝑙𝑖𝑡𝑦 𝑃!" 𝐷!",! 𝑤𝑖𝑡ℎ 𝑃𝑟𝑜𝑏𝑎𝑏𝑙𝑖𝑡𝑦 1 − 𝑃!" 𝐼!" 𝑡 =∝!" 𝐼!" 𝑡 − 1
∝!" =
𝐷!",! 𝑤𝑖𝑡ℎ 𝑃𝑟𝑜𝑏𝑎𝑏𝑙𝑖𝑡𝑦 𝑃!" 𝐷!",! 𝑤𝑖𝑡ℎ 𝑃𝑟𝑜𝑏𝑎𝑏𝑙𝑖𝑡𝑦 1 − 𝑃!"
𝐼!! ’s are the investment costs per unit of PV and battery storage, and ∝!! ’s imply the percentage of decease in investment costs of PV and battery storage.
Model Validation The investment model was built on the basis of a mathematical optimization and the methodologies to calculate periodic cash flows, finance charges and time value of money are
16
adopted from well-established concepts in finance and engineering economics. Micro-grid’s saving is obtained from a two-stage stochastic optimization model, which is convex and a unique optimal solution is guaranteed for it. The functional form that defines micro-grid’s savings is a linear regression model and its accuracy and goodness of fit can be examined with appropriate statistics commonly used in linear regression literature. Therefore, the objective function along with the constraints for the investment model form a convex mixed integer linear or quadratic programming and unique optimal solution is guaranteed for such a problem.
Illustrative Results We expect that the amount and timing of investments on micro-grids will be dependent on economic benefits and capital expenditures. But what is more important is to quantify the impact on investment decisions of capital cost uncertainty and operational stochasticity. To demonstrate this, we solve a single investment problem with and without taking into account the underlying uncertainties. Further analysis will also be performed to demonstrate the sensitivity of investment decisions to the functional form of the micro-grid savings. For the case study, the investment decisions are sought to form a micro-grid over a 4-year time horizon. On-site resources for the micro-grid are to be selected from GF, PV, WT and electricity storage. GF and WT investment costs are considered to be deterministic and known over the course of four years. Gas prices, PV and storage investment costs are assumed to be stochastic with no correlation among them. Electricity price is assumed to be driven by gas price during peak hours. The electricity price for off-peak hours is obtained from a profile explained in [6] and as a percentage peak price. Table 5 lists the parameters that define the dynamics of random variables. Table 5.a -‐ PV Investment Cost 𝐷!",!
0.9
Table 5.b -‐ Storage Investment Cost 𝐷!",! 0.9
𝐷!",!
0.6
𝐷!",!
0.85
𝜎!
0.2
𝑃!"
2/3
𝑃!"
2/3
𝐶!
7 ($/mmBtu)
𝐼!",!
6750000 ($/MW)
𝐼!",!
5200000($/MW)
Table 5.c -‐ Gas Prices 𝛼!
0.045
Given the above input, the possible scenarios realized for each random variable over the course of four years are shown in Figures 3, 4 and 5.
17
6
7
x 10
6
5.5
x 10
18 16
6
5
4
3
2
4
3.5
12 10 8 6
3
1
0
14 4.5
Gas Price $/mmBtu
Storage Investment Cost $/MW
PV Investment Cost $/MW
5
4
0
0.5
1
1.5
2 years
2.5
3
3.5
Figure 3 PV investment cost
4
2.5
0
0.5
1
1.5
2 years
2.5
3
3.5
4
2
0
Figure 4 Stochastic storage investment cost
0.5
1
1.5
2 years
2.5
3
3.5
Figure 5 Stochastic gas prices
6000000 5000000 4000000 3000000 2000000 1000000 0
GF Investment Cost WT Investment Cost Year 1 Year 2 Year 3 Year 4
$/MW
Another set of inputs to the investment model is the capital cost of gas-fired generation unit and wind turbine, which are assumed to be deterministically known for the next four years (see Figure 6).
Figure 6: GF and WT investment cost
Some operational characteristics such as wind turbine efficiency and battery storage duration are fixed to keep the optimization problem computationally tractable. Throughout the illustrative examples we will assume that 𝜂!" = 39% and battery storage duration is 1 hour. There is also a set of financial parameters that are inputs to the investment model (see Table 6):
Table 6 -‐ Financial Parameters 𝐶𝐵! ($)
1,000,000
𝑅𝑂𝐼
2%
𝐹𝐶
4%
𝐹𝑇 (𝑦𝑒𝑎𝑟𝑠)
5
𝐵𝐿𝑖𝑚𝑖𝑡($)
5,000,000
𝐿𝑎𝑛𝑑𝑃𝑟 ($/𝑎𝑐𝑟𝑒𝑠)
20,000
𝐿𝑎𝑛𝑑𝐷𝑒𝑛𝑠
100%
18
4
The last set of inputs refers to the restrictions imposed to invest in on-site resources. These will form the constraints that directly determine the maximum capacity of each resource that can be installed in the micro-grid. Impact of uncertainty is significant For the above micro-grid, a linear regression is built to explain the annual cost of micro-grid conditioned on natural gas prices. The following functional form turns out to be the appropriate fit to the conditional annual cost of micro-grid: 𝐶𝑜𝑠𝑡!",! = 𝛽!,! + 𝛽!,! 𝐼!",! + 𝛽!,! 𝐼!",! + 𝛽!,! 𝐼!",! + 𝛽!,! 𝐼!",! For example, for gas prices at 3.48 $/mmBtu, the coefficients are as shown in Table 7:
Table 7 - Coefficients of micro-grid's cost function GP=3.48 Coefficients Intercept 10024872.8 𝐼!"
-‐597642.28
𝐼!"
-‐6850356.5
𝐼!"
-‐9196434.9
𝐼!"
-‐920318.47
Expected optimal incremental investments over the course of four years are plotted in Figure 7. More investment in wind turbine is due to the fact that its contribution to micro-grid’s savings is more than that of the other resources. Moreover, due to the interdependency between gas and electricity prices, the savings from micro-grid increases when gas prices increase. This would lead to more investment in on-site resources once the gas prices grow higher. Figure 8 depicts the financial activities over the investment horizon. The decision on whether to use own cash or borrowed fund for resource procurement depends on rate of return on invested cash and finance charge. The expected cash flow at the end of horizon (including beyond-horizon projected cash flow) along with its standard deviation is shown in Figure 9. High volatility of cash flow is due to high variance of PV investment cost and gas prices.
19
MW
5 4
E[GF]
3
E[PV]
2
E[WT]
1
E[ST]
0 Year 1 Year 2 Year 3 Year 4
$ Millions
Figure 7: Optimal incremental capacity investments over 4 years; averaged over all scenarios
14
Other Investment
12
Return on Other Investment
10 8
Cash Spent
6 4
Borrowed Fund
2 0 -‐2
Year 1 Year 2 Year 3 Year 4
Cash Flow
$
Millions
Figure 8: Average financial activities over 4 years average over all scenarios
12 10 8 6 4 2 0 E[CF(4)]
SD[CF(4)]
Figure 9: Distribution of cash flow at the end of horizon averaged over all scenarios
20
6
8.6
5.5
x 10
PV Storage
8.4
5
Investment Cost $/MW
Gas price $/mmBtu
8.2
8
7.8
7.6
4
3.5
3
7.4
7.2
4.5
1
1.5
2
2.5 Years
3
3.5
2.5
4
Figure 10: Expected gas prices over investment horizon
1
1.5
2
2.5 Years
3
3.5
4
Figure 11: Expected Photovoltaic and Storage investment costs over investment horizon
Next, we examine the results if we did not consider the uncertainty and represented the random values with their respective expected values. Figures 10 and 11 graph the expected value of gas prices, and storage investment costs over four years of investment horizon.
5
MW
4
E[GF]
3
E[PV]
2 1 0
Millions
Optimal incremental investment decisions are shown in Figure 12. While the decisions on average are not significantly different, considering all stochastic scenarios leads to more distributed investment over 4 years. This would be a better strategy in the presence of uncertainty. Accumulated cash flow at the end of the horizon (including projected future cash flows) is also slightly different compared to the case when uncertainties are considered (Figure 13). 15 10
E[WT]
5
E[ST]
0
Year 1 Year 2 Year 3 Year 4 Feigur 12: Optimal incremental investment over 4 years; deterministic case
Other Investment Return on Other Investment Cash Spent Year 1 Year 2 Year 3 Year 4
-‐5
Borrowed Fund
Figure 13: Financial activities over 4 years; deterministic
As shown in Figure 9, the standard deviation of cash flow at the end of the horizon is significant. The impact of uncertainty will significantly increase as the variations of random variables increase. Suppose that in the above case study we increase the variance in Photovoltaic capital cost. The results are shown in Figures 14-16. Comparing these figures to Figures 7 – 9 shows significant changes in investment strategy and financial activities over 4 years.
21
5
MW
4
E[GF]
3
E[PV]
2
E[WT]
1
E[ST]
0 Year 1 Year 2 Year 3 Year 4
$
Millions
Figure 14: Optimal incremental capacity investments over 4 years; averaged over all scenarios
14
Other Investment
12 10 8
Return on Other Investment
6
Cash Spent
4 Borrowed Fund
2 0 -‐2
Year 1 Year 2 Year 3 Year 4
Cash Flow
$
Millions
Figure 15: Average financial activities over 4 years; averaged over all scenarios
12 11.5 11 10.5 10 9.5 E[CF(4)]
SD[CF(4)]
Figure 16: Distribution of cash flow at the end of horizon average over all scenarios
Different Operational Forms We now demonstrate how investment decisions change as the contribution of each resource to micro-grid’s savings changes. Two different functional forms are examined: 1) savings with both linear and interaction terms, and 2) savings with only interaction terms. The results are shown for a sample path with gas prices shown in Table 8: Table 8 - Sample natural gas prices over 4 years Year
1
2
3
4
22
GP ($/mmBtu)
8.77
10.98
13.75
17.22
$/MW
Millions
Investment costs for various resources are shown in Figure 17. 6 5 GF
4 3
PV
2
WT
1
BS
0 Year 1
Year 2
Year 3
Year 4
Figure 17: Capital cost for resources over 4 years
The first functional form is an example of a micro-grid where the contribution of GF unit to the savings is linear and the savings from the other resources are significant only if we consider the following interactions (coefficients are shown in Figure 18):
Millions
𝐶𝑜𝑠𝑡!",! = 𝛽!,! + 𝛽!,! 𝐼!",! + 𝛽!,! 𝐼!",! ×𝐼!!,! + 𝛽!,! 𝐼!",! ×𝐼!",! 200
No MG Cost
150
GF Cost Reducdon
100 50
PVxWT Cost Reducdon
0
WTxST Cost Reducdon
Year 1 Year 2 Year 3 Year 4
Figure 18 Coefficients of micro-grid's cost function over 4 years
The incremental investment decisions in various resources are shown in Figure 19. The interaction between WT and storage leads to simultaneous investment of these two in year 3. WT dominates the investment because of the higher contribution of this resource to the savings. Once the value of storage in micro-grid is increased by expanding its application, it can becomes more attractive for investment. Having an interaction term between WT and storage can represent a case where storage is not only used for time arbitrage but also coupled with WT’s production.
23
MW
5 4 3 2 1 0
E[GF] E[PV] E[WT] 1
2
3
4
E[ST]
Years Figure 19 Incremental investment decisions over 4 years
In the second example, we assume that the interactions terms are significant and all the other terms are statistically insignificant, i.e., 𝐶𝑜𝑠𝑡!",! = 𝛽!,! + 𝛽!,! 𝐼!",! ×𝐼!",! + 𝛽!,! 𝐼!",! ×𝐼!",! + 𝛽!,! 𝐼!",! ×𝐼!",! + 𝛽!,! 𝐼!",! ×𝐼!",! Incremental investment in resources is shown in Figure 20. Investment in PV and storage are now more significant since these two resources have higher contributions toward micro-grid savings. 5
MW
4 3
E[GF]
2
E[PV]
1
E[WT]
0
E[ST] 1
2
3
4
Years Figure 20 Incremental investment decisions over 4 years
Conclusion In this paper an investment model was developed for micro-grid’s portfolio optimization. The return on investment in a micro-grid is highly dependent on the operation of various on-site resources coupled with the interchange with the utility grid. In this work, the loop between the investment and the operation model was closed using operation models developed by Farzan et al [7]. We also considered two types of uncertainty in investment decisions: 1) short-term or operational variations were built into optimal operation models, and 2) long-term uncertainty associated with investment was addressed by solving the investment model for all possible stochastic scenarios that could be realized over the investment horizon. It was shown that the impact of uncertainty is significant and an investment strategy which is more distributed over the horizon could be a better alternative in hedging against the uncertainty in the future.
24
In addition, different functional forms for micro-grid’s cost (or savings) were examined. The results show that enhancing the application of more expensive resources, such as battery storage, can lead to more savings associated with those resources. Therefore, investment would become more attractive in such resources.
References 1. Asano, H., Ariki, W. and Bando, S. (2010) Value of investment in a microgrid under uncertainty in the fuel price. Power and Energy Society General meeting, IEEE, pp 1-5. 2. Bastian-Pinto, Brandao L., Ozorio L. (2012) A symmetrical binomial lattice approach, for modeling generic one factor Markov processes. Real Options Analysis Workshop, London School of Business, London. 3. Bruno, S. and Sagastizabal, C. (2008) Optimization of Real Asset Portfolio using a Coherent Risk Measure: Application to Oil and Energy Industries, International Conference on Engineering Optimization, Rio de Janeiro, Brazil. 4. Cox, J. C., Ross, S. A., Rubinstein, M. (1979) Option Pricing: A simplified approach. Journal of Financial Economics, No. 7, pp. 229-263. 5. El-Khattam, W., Bhattacharya, K., Hegazy, Y., and Salama, M. M. A. (2004) Optimal investment planning for distributed generation in a competitive electricity market. IIIE transactions on power systmens, Vol. 19, Issue 3, pp. 1674-1684. 6. Farzan, F. and Jafari M. A. (2013) A Real Option Model of Micro-grid Investment under Uncertainty. Working paper, Department of Industrial & Systems Engineering, Rutgers University, NJ, 2013. 7. Farzan, F., Jafari M. A, Masiello, R., Lu, Y. (2013) Towards Optimal Planning and Operation Control of Micro-grids under Uncertainty. Working paper, Department of Industrial & Systems Engineering, Rutgers University, NJ. 8. https://analysis.nrel.gov/homer/. 9. Nelson, D., B., Ramaswany, K. (1990) Simple binomial processes as diffusion approximations in financial model. The Review of Financial Studies, pp. 393-430. 7. 10. Park, C. S., Sharp-Bette, G. P. (1990) Advanced Engineering Economics. John Wiley & Sons Inc. 11. Schultz, R, Tiedemann, S. (2003) Risk Aversion via Excess Probabilities in Stochastic Programs with Mixed-Integer Recourse. SIAM Journal on Optimization, Vol, 14, pp 115 138. 12. Weingartner H. M. (1963) Mathematical programming and the analysis of capital budgeting problems. Princeton-Hall, Englewood Cliffs, NJ. 13. Northern Power® 100 Wind Turbine General Specifications, pp6.
25
