A Tool for Real Time Demand Estimation

5 downloads 70 Views 557KB Size Report
Feb 21, 2007 - With the implementation of availability based tariff. (ABT) - the commercial mechanism [2], from 1st Jan, 2003 and open access regulation [3] in ...
A Tool for Real Time Demand Estimation V.K.Agrawal, IEEE member

U.K.Verma, IEEE member

Abhimanyu Gartia

Rakesh

Abstract: Electrical load forecasting in power system is a most basic requirement to facilitate the generation scheduling and is an important task for ensuring reliability and economic operation. All around the globe a large number of load forecasting procedures / tools are being used and most of these basically use the statistical data to project the demand on daily/ weekly/ monthly/ yearly basis. However, due to the fast changing scenario in real time power system, the SCADA data of the recent past period can also immensely help the operator towards demand side management for the immediate future period say for next half an hour to one hour more accurately and quickly. In India with the increase in the size of the grids as well as large interconnections, along with the increased complexities towards power system operation and control, there has also been a phenomenal expansion and growth of power markets. These markets have assumed greater importance, particularly after the advent of The Electricity Act 2003, introduction of open access in transmission and emergence of power exchanges. In this paper an attempt has been made to deliberate on an approach which is helping in estimating the demand more accurately for shorter durations on a real time basis. The basic principle adopted in this approach is that it applies the simple time series analysis to estimate the demand in the dynamically changing grid conditions. This tool is in practical use and has been found to be of immense benefit to the grid operators at various state and regional load dispatch centres (LDC) in the country. With this utility the operators are able to estimate the demand for future 30 minutes on a real time basis and this is helping them in taking an advance action towards ensuring the grid operation with reliability, security and economy. Key words: Demand estimation, Frequency estimation, Dynamic correction, Real time operation,

With the implementation of availability based tariff (ABT) - the commercial mechanism [2], from 1st Jan, 2003 and open access regulation [3] in the year 2005, an all round reform and competition has been brought out in electric supply industry in SR leading to better services to the end-user. However, all these changes have also brought additional complexities in power system operation as well as increase in the transmission elements loading. This at times also leads to

Demand

Frequency

23000

50.5

22000 50

21000 20000

49.5

19000 49

18000 10:00PM

8:00PM

6:00PM

4:00PM

2:00PM

Tim e (In Hrs)

Fig. 1: SR Demand-profile on 21st Feb, 2007

In line with the general practice followed world wide, the unified load dispatch and communication schemes implemented at different control centers in India during 200204 are equipped with load forecasting tools to assist the grid operator. These tools primarily make use of the hourly data and the algorithm is generally based on similar day forecast or weather based forecast. Due to the use of hourly sampling of data, both in history and forecast, forecasting for the smaller time interval within an hour poses one of the major constraints in these methods and in this paper attempt has been made to address this issue. This utility is aimed to accurately estimate grid demand and frequency on real time basis for a shorter

978-1-4244-4331-4/09/$25.00 ©2009 IEEE __________________________________________________________ # Data pertains to 2007 conditions, during implementation of algorithm

12:00PM

10:00AM

8:00AM

6:00AM

48.5 4:00AM

17000 2:00AM

Since in the SR demand large chunk of load comes from the agricultural sector, the weather in region has a telling effect on the demand profile. The region generally experiences its peak demand in the months of February-March every year, which tapers down from April onward and goes to the minimum level during peak monsoon month of July / August. In addition to large variation in demand profile (in the range 3000MW to 4000MW) from peak to off-peak period, appreciable frequency variation is also observed in SR in a day as shown in fig. 1.

Typical SR Demand with Frequency

12:00AM

Indian Southern regional (SR) grid is a large power system having an installed capacity of about 37,000 MW, meeting a peak demand of over 24,000 MW and covering a geographical area of approx. 6,51,000 sq.km. It comprises of electrical system of 4 States viz. Andhara Pradesh, Kanataka, Kerala, Tamil Nadu and Union Territory of Puducherry. The consumer profile in Southern Region is predominantly industrial in nature, followed by agricultural and domestic load. The break-up percentage of different categories of load is [1]: industry 37.25%, agriculture- 30.10%, domestic - 22.51%, commercial - 6.34%, others 3.80%#.

encroachment of the security margins and hence demands more surveillance. These changes have therefore necessitated adaptation of best possible grid side management techniques including use of better load forecasting methods so that with the prior estimation of load / demand, proactive action is taken to avoid any contingency.

Dem and (In MW)

1. Introduction

1

duration future time window say for next immediate 30 minutes. In order to achieve this with a high level of accuracy, well defined and established algorithms like similar day forecast (SDF) [9] associated with statistical time series analysis have been used on historical and current demand profile with dynamic correction. The paper discusses the tool developed and implemented at southern regional load dispatch centre (SRLDC) Bangalore and some other LDCs in India, its performance and results based on last two years experience and scope for further improvement. 2. Provision of load forecasting in India Indian electricity grid code (IEGC) [8] order issued by Central Electricity Regulatory Commission (CERC) has given due importance to load forecasting for operational purposes and Clause 5.3.3 of IEGC is quoted below: Quote: “Each State/SLDC shall develop methodologies/mechanisms for daily/weekly/monthly/yearly demand estimation (MW, MVAR and MWH) for operation purposes. The data for the estimation shall also include load shedding, power cuts etc. SLDC’s shall also maintain historical data base for demand estimation.” Unquote: 2.1 Practice being followed World wide The response of the power utilities worldwide with regard to load forecast exercise reveals a wide spectrum of approaches. The methods range relatively simple procedure to a complex algorithm. References of some of the approaches which have been tried over the years and improved with time are given below: 2.3 New South Wales Electricity System-Australia [4] They have used Bayesian semi parametric regression methodology to model intra-day load data and forecast the demand. They have taken daily periodic, weekly periodic load and temperature sensitive load component as one of the inputs for demand forecast.

2.6 Power System IEEE Transactions 1993 [7] In this transaction the approach is based on generalized algorithm combining knowledge based and statistical techniques. The generalized model is used for weather – load relationship. The algorithm uses pair wise comparison to qualify categorical variables and utilizes regression to obtain least square estimated load forecast. In most of the referred cases, projected demand for shorter durations on real time basis is being addressed to a certain extent by extrapolation of past data using regression analysis. 3. Proposed method of demand estimation The proposed method developed at SRLDC addresses the issue of load forecasting for short duration for less than an hour using hybrid algorithm comprising SDF technique [9] and statistical co-relation of load forecast data. The dynamic correction has also been incorporated using real time values obtained from supervisory control and data acquisition (SCADA). In the method, the time window for demand estimation has been kept for 30 minutes but the same is configurable on real time basis. It has been proposed for 1 day data analysis with 1440 samples (one sample for each minute) where as in the conventional method for short term demand forecast, 7 days with 168 samples (one sample for each hour) is considered for analysis purpose. The demand estimation so obtained is also targeted for real time correction. The algorithm is arrived after detailed analysis of the load forecast practices followed in various utilities referred above and the specific requirement in Indian context. In order to carry out demand estimation for short duration, actual demand data profile during a certain month say February 2007 has been mathematically analyzed for a statistical co-relation with previous days historical demand profiles. Typical comparison of demand data for a typical day say 7th Feb 2007 has been shown in Fig. 2

Typical Demand Variation 31st Jan

2.4 National Grid Company UK [5]

6th Feb

7th Feb

24000

22000 21000 20000

They have used dynamics non-linear model consisting of multiple time series analysis and forecasting. Also some utilities in Brazil are using the extrapolation of past data or regression. Here they are using the real time data for the forecast.

10:00PM

8:00PM

6:00PM

4:00PM

10:00AM

8:00AM

6:00AM

4:00AM

2:00AM

12:00AM

18000

2.5 Industrial Electricity computation in Brazil [6]

2:00PM

19000

12:00PM

Demand (MW) --->

23000

Short term forecasting modeling involves precise customer information, better weather prediction and major load dependent events like TV programmes etc. The best results obtained in UK are having an error of 2% within any half-hour period.

Time (In Hrs) --->

Fig. 2 Typical demand variation in SR

978-1-4244-4331-4/09/$25.00 ©2009 IEEE

2

On scrutiny of the above data it can be observed that compared to the last week’s data the data for the previous day is better matched. The differential demand between 6th and 7th Feb 2007 is mostly positive (+ve) and has a standard deviation of 434 MW. The pattern is generally valid for other days also. Further, it has also been observed that the historical reference profile for a working day (Monday to Saturday) is best matched with the previous working day, whereas for Holiday (Sunday and Closed Holiday), it is best matched with the previous holiday. Based on this observation, the load profile of the previous day and/or previous holiday has been taken as the historical reference profile. In economic term, all developing countries have similar characteristic in public utilities like power, water supply, traffic movement and etc. The SDF algorithm is extended to each data point of 1 minute interval to include detail demand variations in shorter durations as compared to conventional 1-hour interval data points. The proposed algorithm flow chart is given in fig. 3.

 A and B are the statistical constants to be derived on the basis of the forecasted demand vis-à-vis actual demand and worked out in the following manner [7]:

START

Request SCADA registration NO

YES Forecast demand profile data prepared at the start of the day

Update real time demand and frequency data from SCADA up to current time

Error message, migrate to backup server

Compute delta error on forecast profile and demand from 00hrs to current time

The process of demand estimation made in two steps using the proposed algorithm is as under: 3.1 Profile preparation

Compute statistical constant for delta error adapting linear to complex equation [10]

Initially the forecasted demand for the whole of the next day at 1 minute interval is worked out as given below [9]: (Fmax - Fmin) Fi = Fmin + ------------------ * (Hi – Hmin) (1) (Hmax – Hmin) Where Fi=Initial Forecasted Demand for the ith minute with i varying from 1 to 1440. Fmax= Forecasted Maximum Demand Fmin= Forecasted Minimum Demand Hi = Historical Demand for the ith minuite. Hmax = Historical Maximum Demand Hmin = Historical Minimum Demand

Extrapolate the delta error for future 30 minute using the statistical constants

Estimate demand for future 30 minute using forecast profile and delta error

Estimate frequency for future 30 minute using current frequency and delta demand

Here, the Forecasted Minimum and Maximum demand are user point and depends upon the specific power transaction made and generation availability for the day.

Update the results to SCADA for user displays, historical trends and other applications

3.2 Demand estimation and dynamic correction During real time operation, the forecasted demand is continuously moderated for next 30 samples (at an interval of 1 minute each) by extrapolating it on the basis of the difference observed between the historical demand and actual demand met from 00:00 Hrs upto the point of estimation, using the following linear equation: y(jth) = A + B * x(jth)

(2)

where,  y(jth) is the correction to be made to the earlier forecasted demand for next 30 samples, i.e, next 30 minutes (j= 1 to 30)  x(jth) is the sample number (varies from 1 to 30)

END

Fig. 3 Flow chart for demand and frequency estimation    

[ n ∑ p.q – ∑ p ∑q ] ---------------------------[ n ∑p2 – (∑p)2 ] A = qavg.– B * pavg. p (ith) = ith sample no.(varies from 1 to 1440) q (ith)=[D(ith) – F(ith)]

B =

978-1-4244-4331-4/09/$25.00 ©2009 IEEE

(3) (4)

3

where q(ith) is the differential demand between actual demand D(i) and initial forecasted demand F(i) for ith minute (varies from 1 to 1440). From (1) and (2), the estimated demand for next 30 minute is moderated as given below: (5) F(jth) = F(i) + y(jth) Where F(jth ) = Final Forecasted Demand F(i) = Initial Forecasted Demand y(jth) = Differential demand estimation Demand estimation is done for a short duration of 30 minutes rolling window. Analysis of estimated samples reveals that as we move away towards future time scale, there is a fall in accuracy due to inclusion of more uncertainty. Considering the response of grid control device and the methods of passing instruction on a hierarchical operation in India, the 30 minute time scale has been fixed which is sufficient for grid operator to take advance action for possible correction. All computations using the above algorithm are carried out on SCADA real time data for giving real time demand estimation. 3.3 Frequency estimation From grid operation and security point of view, frequency is very vital parameter. At every point in real time the estimated frequency can be computed to reasonable accuracy by the formula given below [ F(jth ) – Da ] Fe = Fa – ---------------- * PN (4) Da Where Fe = Estimated frequency Fa= Actual Frequency in real time at the time of Computation F(jth) = Estimated demand Da = Actual Demand at the time of computation PN = Power Number of the system derived from the system data comprising generation composition and load pattern. Typical value varies in the range 2%-5%.

The frequency estimation done by above formula, does not take into account the Generation added / removed from grid during the rolling estimation time window. In order to account for such variation in the Generation, provision has been made for user entry as below: [ F(jth ) – Da ) + Ga ] Fe = Fa – -------------------------- * PN (7) Da Where Ga = Generation correction (added/removed) 4. Case study The proposed method has been implemented in actual practice at SRLDC Bangalore from February 2007. As explained in (1), minute wise demand sample for a typical day say 8th February 2007 is taken as reference profile for demand estimation on the next day say 9th February 2007. Using the algorithm in (2) to (7), the estimated demand, estimated frequency and the expected maximum change in demand and frequency in the future 30 minutes in a rolling window are calculated and displayed on the SCADA monitor on a real time basis as shown in fig. 5 along with current day actual demand, the estimated demand comparison. Also the comparison of estimated demand and actual demand for a complete day for the selected day i.e for 9th February 2007 is depicted in Fig. 4 which indicates a close matching pattern between actual demand and estimated demand. The algorithm adopted for estimating demand gives automatic correction, if during this period some corrective action is taken by any system / sub-system. So there is a continuous self-correcting mechanism built in the algorithm resulting in good accuracy. This feature has been demonstrated by selecting an actual incident that took place on 9th February, 2007 when at around 14:10 Hrs. (i.e. 850th sample) one unit of 210 MW at Raichur Thermal Power Station in Karnataka tripped and the consequent error in demand estimation has been automatically corrected in next 5 minutes approx. as clearly depicted in Fig. 5. On comparing the actual demand met and estimated demand, the maximum variation observed is +/–100 MW with average variation even less than 30MW on a 22000MW demand scale and hence a satisfactory level of accuracy could be achieved as shown in Fig. 6. Actual Vs Estimated SR Demand For 9th Feb,2007 Estimated

22000 21000 20000 19000

10:00PM

8:00PM

6:00PM

4:00PM

2:00PM

12:00PM

10:00AM

8:00AM

6:00AM

4:00AM

2:00AM

18000 17000 12:00AM

Demand (MW)--->

Actual 24000 23000

Tim e (In Hrs)--->

Fig. 4 SCADA displays with estimated demand and frequency during changeovers

Fig. 5 Comparison of actual and estimated demand in a day

978-1-4244-4331-4/09/$25.00 ©2009 IEEE

4

Also, the comparison between actual demand met and estimated demand over a typical period of 30 days from 8th February - 9th March, 2007 have been done and the same is shown in Fig. 7. This clearly indicates that the average variation is consistent and is in the range of +/- 40 MW. The algorithm function thus used has been quite useful for advance warning specifically during load change over, peak time, off peak time and other times of demand variations as demonstrated in Fig. 5. Actual & Estimated Difference For 9th Feb,2007 Demand (In MW)

600 400 200 0 -200 -400 10:00PM

8:00PM

6:00PM

4:00PM

2:00PM

12:00PM

10:00AM

8:00AM

6:00AM

4:00AM

2:00AM

12:00AM

-600

demand due to load dependent events such as Cricket Match, etc requiring necessary correction of the projected demand over short duration is also being studied further. 5. Conclusion The development is a culmination of extensive studies of power system operation in India and abroad on the uncertainty that always prevails. The algorithm and implementation is in practical use in few places in India like SRLDC Bangalore, LDC Chennai, LDC Kalamessary and LDC Jabalpur. The purpose of short duration demand and frequency estimation is to assist grid operators on real time operations. It has helped grid operators in the time of crisis. The methodology can be further extended to other few important grid parameters like Power flows, Load angle, Bus voltage etc in power systems. All the public utilities experiencing similar response, the algorithm can be further analyzed for case studies in other areas like water supply, traffic movement etc.

Tim e (In Hrs)

Fig. 6 Typical actual vs estimated demand difference in a day

References 16th Electric Power Survey Report released by Central Electricity Authority - January 2001. 2. Availability based tariff implementation in western region http://www.cercind.gov.in/290702/86-02.pdf 3. Grant of open access http://www.cercind.gov.in/23032005/198-04.pdf 4. Modeling & Short-term Forecasting of New South Wales Electricity System Load by Michael Smith (NSW 2006). 5. Load Forecasting practices in NGC,UK by Anne Ku (Global Energy Business,2002/Mar-Apr) 6. Short-Term Forecasting of Industrial Electricity consumption in Brazil by Regina Sadowink & Emanuel Pimentel. 7. A generalized knowledge-based short-term loadforecasting technique by Rahman.S.Hazim,O.Bradley IEEE Transactions,May 1993. 8. http://www.wrldc.com/docs/gridcode.pdf 9. XA21 SDF Functional Design guide by GE Harris USA 10. Research Design and Statistical Analysis By Jerome L myers and Arnold D Well (http://books.google.com). 1.

Estimated Vs. Actual Variation

Deviation(M W)--->

Pos. Deviation

Neg. Deviation

100 50 0 -50 -100 0

5

10

15

20

25

Days--->

Fig. 7 Actual vs estimated demand differences over a month 4.1 Analysis of extreme cases Since the computation period is 24 hrs, at the beginning of the day i.e. at 00: 00 hrs, limited number of current demand sample data is available for statistical analysis with the historical demand, Therefore at the beginning of the day the estimation may loose the accuracy for some time as compared to a point where sufficient historical data are available, say any time after 01:00 Hrs. It is felt that the problem can be resolved by taking a wider window (say two days) of demand profile for history as well as actual demand met. This may lead to a large calculation points for statistical computation and hence consequent loading on the real time system, which needs to be further analyzed. . On sudden variation of demand due to forced outage or load crash in any system/sub-system of large magnitude say above 20% of demand being met, the inter-dependency of the demand profile may be severely affected for a short duration, which ultimately would die out as per the statistical algorithm but it may take some more time. Similarly sudden change in

Authors Mr. V.K. Agrawal is presently working as General Manager at NRLDC, New Delhi and is responsible for operation of the power system grid in Northern Region and for maintaining its safety and security. Mr Agrawal has done his graduation in Electrical Engineering in 1977 from Delhi College of Engineering and M.Tech in Power Apparatus and Systems from I.I.T., Delhi. He has a long experience in respect of Regional Grids Operation and Control and work on associated Commercial aspects.

978-1-4244-4331-4/09/$25.00 ©2009 IEEE

5

Mr UK Verma is presently working as Additional General Manager at Southern Regional Load Despatch Centre Bangalore. He did his graduation in Electrical Engineering from Birla Institute of Technology Ranchi in 1980. He is a member of IEEE. He has worked in NHPC Ltd. and has rich experience in O&M of Hydro Power Plant. He is presently responsible for the operation of Southern Region Electrical Grid.

Mr Abhimanyu Gartia has done post graduation in Electrical Power System from Indian Institute of Technology Kharagpur in 1986 is presently working as Chief Manager at Southern Regional Load Despatch centre Bangalore. He has worked for the unified load dispatch schemes for southern and western region in India. His field of interests are SCADA and power system stability.

Mr. Rakesh has done graduation in Electronics Engg. From National Institute of Technology, Surat in 2001. He is presently working as Deputy Manager at Southern Regional Load Despatch centre Bangalore. His field of interests are SCADA and web technologies.

978-1-4244-4331-4/09/$25.00 ©2009 IEEE

6

Suggest Documents