Chapter 20
Improving WRF GHI Forecasts with Model Output Statistics Burak Barutcu, Seyda Tilev Tanriover, Serim Sakarya, Selahattin Incecik, F. Mert Sayinta, Erhan Caliskan, Abdullah Kahraman, Bulent Aksoy, Ceyhan Kahya, and Sema Topcu
Abstract Solar energy applications need reliable forecasting of solar irradiance. In this study, we present an assessment of a short-term global horizontal irradiance forecasting system based on Advanced Research Weather Research and Forecasting (WRF-ARW) meteorological model and neural networks as a post-processing method to improve the skill of the system in a highly favorable location for the utilization of solar power in Turkey. The WRF model was used to produce 1 month of 3 days ahead solar irradiance forecasts covering Southeastern Anatolia of Turkey with a horizontal resolution of 4 km. Single-input single-output (SISO) and multi-input single-output (MISO) artificial neural networks (ANN) were used. Furthermore, the overall results of the forecasting system were evaluated by means of statistical indicators: mean bias error, relative mean bias error, root mean square error, and relative root mean square error. The MISO ANN gives better results than the SISO ANN in terms of improving the model predictions, provided by WRF-ARW simulations for August 2011. Keywords Model output statistics • Artificial neural networks • Weather research and forecasting • Turkey
B. Barutcu (*) Istanbul Technical University Energy Institute, ITU Ayazaga Campus, Maslak, Istanbul 34469, Turkey e-mail:
[email protected] S.T. Tanriover • S. Sakarya • S. Incecik • F.M. Sayinta • E. Caliskan A. Kahraman • C. Kahya • S. Topcu Department of Meteorology, Istanbul Technical University, Maslak, Istanbul, Turkey B. Aksoy Turkish State Meteorological Service, Ankara, Turkey © Springer International Publishing Switzerland 2015 I. Dincer et al. (eds.), Progress in Clean Energy, Volume 1, DOI 10.1007/978-3-319-16709-1_20
291
292
20.1
B. Barutcu et al.
Introduction
Short-term irradiance forecasting is an important issue in the field of solar energy for many applications. Numerical weather prediction (NWP) models have proven to be powerful tools for solar radiation forecasting. The use of numerical meteorological models in combination with statistical post-processing tools may have the potential to satisfy the requirements for up to 72 h ahead of irradiance forecast. In this study, we present an assessment of a short-term irradiance system based on the Advanced Research WRF (WRF-ARW) (V3) meteorological model and a postprocessing method in order to improve the overall skills of the system for a full month simulation of the August 2011 over the Southeastern Anatolia Region (SEA) of Turkey.
20.2
Data and Methodology
20.2.1 Data The study is carried out in SEA which is a highly favorable location for the utilization of solar power in Turkey (Fig. 20.1). SEA is located between the latitudes of 36 –38 N, covering about 7.5 million hectares of land which is approximately 10 % of the whole country. Global horizontal solar irradiation (GHI) data
Fig. 20.1 GHI measurement sites in Southeastern Anatolia Region of Turkey
20
Improving WRF GHI Forecasts with Model Output Statistics
293
Table 20.1 Geographical elements of GHI measurement sites in Southeastern Anatolia Region Site
Latitude ( N)
Longitude ( E)
Altitude (m)
Bozova Ceylanpınar Kilis Mardin S¸ırnak
37.37 36.84 36.71 37.31 37.52
38.51 40.03 37.11 40.73 42.45
622 360 640 1,040 1,350
Table 20.2 Parameterizations used in WRF model for the August 2011 simulations Parameterization type
Scheme
Surface layer
Monin-Obukhov with Carlson-Boland (MM5 similarity) Noah land surface model Yonsei University scheme RRTMG Kain-Fritsch (for domains 1 and 2, none for the innermost domains) New Thompson scheme
Land surface model Planetary boundary layer model Atmospheric radiation Cumulus Microphysics
were collected from five meteorological stations, representing the region of interest, which, namely, are Bozova, Ceylanpınar, Kilis, Mardin, and S¸ırnak. To measure the GHI, Kipp & Zonen type pyranometers are used by the Turkish State Meteorological Service. The hourly observational GHI data from SEA are used for model verification. The stations are distributed over a complex terrain with station elevations ranging from 360 to 1,350 m, above sea level (Table 20.1). The full month of August 2011 is selected for the present work.
20.2.2 The WRF Model The Weather Research and Forecasting (WRF) mesoscale meteorological model is applied to calculate the GHI over the study area. WRF is an Eulerian, 3-dimensional, non-hydrostatic mesoscale model with state-of-the-art parameterizations (radiation, planetary boundary layer, surface layer, land surface model, cumulus, and microphysics) in a massively parallel computing environment. Table 20.2 presents the parameterizations used in the WRF model run for this study. In this study, the WRF model concerning the period of August 2011 was run on an initial regional grid of 36 36 km in space covering Europe. Inner domains following with 12 12 km and finally 4 4 km of horizontal resolution with 35 vertical layers over SEA of Turkey were used. The National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS) data with 6 h temporal and 1 spatial resolution were used for initial and boundary conditions.
294
B. Barutcu et al.
The GHI forecasts up to 72 h were performed using the WRF model. The conducted simulations were initialized from 00UTC of each day of August 2011. Two-way nesting option was selected for all nesting operations. It is known that the mesoscale meteorological models can predict GHI without mean bias error (MBE) for clear sky conditions only. However, the bias was highly dependent on cloudy conditions and becomes strong in overcast conditions. Remund et al. [1] evaluated different NWP-based GHI forecasts in the USA, reporting relative root mean square error (rRMSE) values ranging from 20 to 40 % for a 24 h forecast horizon. Similar results were reported by Perez et al. [2], evaluating NWP-based irradiance forecasts in several places in the USA. Remund et al. [1] examined NWP biases compared to a single site and found that the next day (24 h), GHI forecasts of European Centre for Medium-Range Weather Forecast (ECMWF), a NWP model, and GFS have some MBE of 19 %. This MBE was found to be approximately constant for intraday (hour ahead) to 3 days ahead forecast horizons. Lorenz et al. [3, 4] evaluated several NWP-based GHI forecasts in Europe, using statistical measures such as RMSE and MBE. In the chapter by Lorenz et al. [3], results showed rRMSE values of about 40 % for Central Europe and 30 % for Spain. Evaluating ECMWF’s accuracy in Germany, Lorenz et al. [4] also showed that NWP MBE was largest for cloudy conditions with moderate clear sky indices, while forecasted clear conditions were relatively unbiased. They reported rRMSE values of about 35 % for single stations for 24 h horizon forecasts.
20.2.3 MOS and ANN Post-processing techniques are usually applied to refine and improve the NWP model outputs. Model output statistics (MOS) is a post-processing technique used to objectively interpret numerical model output and produce site-specific forecasts. MOS relates observed meteorological parameters to appropriate variables (predictors) via a statistical approach. Hence, they can be used to reduce systematical forecast errors. There are several papers in literature about the MOS correction of GHI forecasts from NWP models [2, 5–7]. They indicate that the MOS application to the WRF GHI forecasts was successful in minimizing bias and reducing RMSE. Bofinger and Heilscher [7] used MOS locally with ECMWF GHI forecasts to create daily solar electricity predictions, accurate to 24.5 % rRMSE for averaged daily forecasts. In this study, WRF model forecasts were used as the predictors. As an alternative to conventional approaches, artificial neural networks (ANNs) have been successfully applied for solar radiation estimation in the literature [5, 6, 8–10]. ANNs recognize patterns in the data and have been successfully applied to solar forecasting. Using training data, ANNs have been developed to reduce rRMSE of the GHI values. The innermost domain was used in the evaluation procedure. Furthermore, a group of post-processes were used: single-input single-output (SISO) ANN and multi-input single-output (MISO) ANN.
20
Improving WRF GHI Forecasts with Model Output Statistics
295
Feed-forward multilayer ANNs consist of three basic components: an input layer, one or more hidden layers, and an output layer. Due to the fact, there isn’t any connection between the nodes in the same layer. The input layer of ANN transfers the information it receives from outside to the internal processing unit. Hidden layers are used for feature extraction during the flow of information. Output layer gives the system output. There are two major steps in the use of a feed-forward ANN, one is teaching and the other is recall [11, 12]. In this study, a group of ANNs were used to reduce the forecast errors. The ANN architectures which are applicable to this goal can be grouped under two main branches. In the first group which is called SISO ANNs, there is one node as input, one neuron as output layer, and a group of (e.g., 5, 7, etc.) neurons are applied as the number of neurons in their hidden layer. In the second group which is called MISO ANNs, a group of nodes (e.g., 3 or more) are set as input layer, and one neuron is applied to the output layer. In this group, embedding technique is used to improve the success, and two different selection criteria are used separately for determining the number of neurons in the hidden layers, one is building as a triangular architecture (n + 1)/2 (where n is the number of nodes in the input layer) and the other is HechtNielsen’s 2n + 1 approximation [13] (which is based on Kolmogorov’s proof [14]). A series of trials were performed to define the embedding dimension of the MISO ANN. Figure 20.2 represents the chosen ANN structure that consists of three nodes as input layer, seven neurons for the hidden layer, and a single neuron as the output layer. In both groups, two third of the signal is set for the training and the rest one third is used for the test. For the sake of comparison, five ANNs were created for each model in both architectures, and these ANNs were trained separately by Levenberg-
Fig. 20.2 The feed-forward MISO ANN architecture used to approximate the WRF outputs to the observation values. The ANN has logarithmic sigmoid activation functions in hidden layers and pure linear activation functions in output layers. Number of hidden layer neurons determined by Hecht-Nielsen’s approximation as 2 3 + 1 ¼ 7 neurons
296
B. Barutcu et al.
Marquardt training algorithm for 500 steps. Logarithmic sigmoid function is used for all hidden layers (because solar irradiance is a nonnegative signal), and pure linear activation function is applied to all output layers. SISO ANNs could never achieve success levels as high as MISO architectures because MISO networks are designed by the embedding approach in prediction theory. Besides, SISOs do not use any information about the past values of a signal.
20.2.4 Adjustment of ANN The best performed neural networks used to improve the WRF results are MISO ANNs. The logarithmic sigmoid activation function was used—as the irradiation data is a nonnegative signal—in the hidden layer. For the output layer, linear activation function proved to produce better results. The performance of ANNs depends on the weights and biases which are selected as small random numbers in the initialization step of ANNs. Therefore, starting with different weights and biases may yield different consequent performances. Keeping this in mind, five separate ANNs were trained for each station and their predictions for 24, 48, and 72 h. To be able to sufficiently diverge from the step until error reduction is no longer possible, all networks were trained for 500 steps with Levenberg-Marquardt algorithm, one of the best among quasi-Newtonian methods. The ANNs with the best performance for each prediction horizon were selected. As a result, 15 ANNs were established for five stations.
20.3
Results
In this study, an assessment of a short-term GHI forecasting system based on WRF GHI forecasts and a post-processing method, ANN for a full month of August 2011, was presented. The WRF model outputs and the ANN post-processes were compared by means of statistical error parameters, as shown in Table 20.3. These aforementioned parameters are as follows: mean bias error (MBE), relative mean bias error (rMBE), root mean square error (RMSE), and relative root mean square error (rRMSE). In order to improve the solar irradiance forecasts, we employed the ANN methodology. The statistical accuracy by the means of presented error values of verification of the WRF outputs to observations is shown in Table 20.3. The impact of ANN on RMSE and MBE or the performances of the ANNs for forecasts of 24, 48, and 72 h selected for this study are presented in Table 20.4. Tables 20.3 and 20.4 as a whole summarize the results for each station separately by showing the full August statistical measures, averaged over the stations for GHI.
20
Improving WRF GHI Forecasts with Model Output Statistics
297
Table 20.3 Error statistics of WRF simulations for August 2011 of Southeastern Anatolia Region in Turkey Forecast horizon
Statistical parameters
WRF verifications Bozova Ceylanpınar
Kilis
Mardin
S¸ırnak
24 h
MBE (W/m2) rMBE RMSE (W/m2) rRMSE MBE (W/m2) rMBE RMSE (W/m2) rRMSE MBE (W/m2) rMBE RMSE (W/m2) rRMSE
26.704 0.094 51.524 0.182 27.349 0.095 51.535 0.179 26.662 0.094 54.239 0.191
32.665 0.113 54.110 0.190 32.101 0.111 51.664 0.179 31.517 0.110 54.242 0.191
31.922 0.107 53.564 0.180 28.556 0.096 55.693 0.187 29.226 0.100 54.391 0.186
34.715 0.116 60.365 0.202 30.694 0.102 56.099 0.187 30.950 0.105 60.603 0.206
48 h
72 h
47.109 0.171 71.995 0.262 46.55 0.169 71.672 0.261 45.814 0.169 71.214 0.264
Table 20.4 Error statistics of ANN post-processes and for August 2011 of Southeastern Anatolia Region in Turkey Forecast horizon 24 h
48 h
72 h
Statistical parameters 2
MBE (W/m ) rMBE RMSE (W/m2) rRMSE MBE (W/m2) rMBE RMSE (W/m2) rRMSE MBE (W/m2) rMBE RMSE (W/m2) rRMSE
WRF + ANN verifications Bozova Ceylanpınar
Kilis
Mardin
S¸ırnak
17.062 0.056 31.407 0.103 19.468 0.064 45.099 0.148 16.808 0.057 39.354 0.133
18.899 0.062 30.808 0.101 22.327 0.073 33.020 0.109 19.896 0.066 33.025 0.110
3.342 0.011 13.153 0.045 3.377 0.012 14.739 0.050 2.591 0.009 16.037 0.055
6.182 0.021 17.938 0.061 9.557 0.023 23.010 0.078 5.089 0.017 29.724 0.100
5.492 0.021 33.912 0.127 6.712 0.025 35.522 0.135 10.143 0.039 37.881 0.146
The following are some regional conclusions based on the comparison of the results given in Tables 20.3 and 20.4: • The regional accuracy of the WRF model outputs in rRMSE for SEA lies in the order of 20.32 % and is reduced to 8.74 % with ANN, for the first 24 h. • Regarding the whole region, the rRMSE is reduced from 19.86 to 10.40 % and from 20.76 to 10.88 % for 48 and 72 h of forecast horizons, respectively. • The RMSE reduction for August via ANN, regarding the 24 h WRF outputs (from 58.31 to 25.44 W/m2), shows 56 % forecast improvement.
298
B. Barutcu et al.
The quality of the forecast is also dependent on the location. The two best predicted sites with the lowest RMSE are places in Mardin and S¸ırnak with more sunny days than the others.
20.4
Concluded Remarks
In this study, The WRF model was used to produce 1 month of 3 days ahead solar irradiance forecasts covering Southeastern Anatolia of Turkey. Additionally, in order to improve the forecasts of GHI, we employed the ANN post-processing. The presented study gives the assessments of short-term GHI forecasting by WRF and a post-processing tool, ANN. The analysis was performed on a full month of August 2011 over SEA of Turkey. The evaluation focuses on the capability of the ANN methodology to improve the forecast of global horizontal solar irradiance by reducing the systematical error. The MISO ANN presents better results than the SISO ANN in terms of improving the model predictions provided by WRF-ARW simulations. Besides, the ANN had a significant performance in August 2011. On the basis of GHI measurements, the present study makes it possible to obtain more accurate predictions. Acknowledgements This paper is funded by TUBITAK COST 111Y234 Project.
References 1. Remund J, Perez R, Lorenz E (2008) Comparison of solar radiation forecasts for the USA. In: Proceedings of the 23rd European PV conference, Valencia, pp 1–9 2. Perez R, Kivalov S, Schlemmer J, Hemker K Jr, Renne´ D, Hoff TE (2010) Validation of short and medium term operational solar radiation forecasts in the US. Sol Energy 84 (12):2161–2172 3. Lorenz E, Remund J, Mu¨ller SC, Traunmu¨ller W, Steinmaurer G, Pozo D, Ruiz-Arias JA, Fanego VL, Ramirez L, Romeo MG, Kurz C, Pomares LM, Guerrero CG (2009) Benchmarking of different approaches to forecast solar irradiance. In: Proceedings of the 24th European photovoltaic solar energy conference, Hamburg, pp 4199–4208 4. Lorenz E, Hurka J, Heinemann D, Beyer HG (2009) Irradiance forecasting for the power prediction of grid-connected photovoltaic systems. IEEE J Sel Topics Appl Earth Observ 2 (1):2–10 5. Rinc on A, Jorba O, Baldasano JM, Delle Monache L (2011) Assessment of short term irradiance forecasting based on post-processing tools applied on WRF meteorological simulations. In: State-of-the-Art Workshop, COST ES 1002: WIRE: Weather Intelligence for Renewable Energies. COST Action WIRE, Paris, 22–23 March 2011, pp 1–9 6. Diagne M, David M, Lauret P, Boland J, Schmutz N (2013) Review of solar irradiance forecasting methods and a proposition for small-scale insular grids. Renew Sustain Energy Rev 27:65–76
20
Improving WRF GHI Forecasts with Model Output Statistics
299
7. Bofinger S, Heilscher G (2006) Solar electricity forecast-approaches and first results. In: 21st European photovoltaic solar energy conference, no. 9, pp 4–8 8. Mathiesen P, Kleissl J (2011) Evaluation of numerical weather prediction for intra-day solar forecasting in the continental United States. Sol Energy 85(5):967–977 9. Rahimikhoob A (2010) Estimating global solar radiation using artificial neural network and air temperature data in a semi-arid environment. Renew Energy 35(9):2131–2135 10. Benghanem M, Mellit A, Alamri SN (2009) ANN-based modelling and estimation of daily global solar radiation data: a case study. Energy Convers Manag 50(7):1644–1655 11. Beale R, Jackson T (1998) Neural computing—an introduction. IOP, Bristol 12. Haykin S (1999) Neural networks: a comprehensive foundation. Prentice Hall, Upper Saddle River 13. Hecht-Nielsen R (1990) Neurocomputing. Addison-Wesley, Reading 14. Kolmogorov AN (1957) On the representation of continuous functions of many variables by superposition of continuous functions of one variable and addition. Dokl Akad Nauk SSSR 114:953–956. Translated in: Amer Math Soc Transl 28:55–59 (1963)