Improving performance of NN based electricity price forecasting using sensitivity analysis

This paper focuses on sensitivity analysis of neural network (NN) parameters in order to improve the performance of NN based short-term electricity price forecasting. Sensitivity analysis of NN parameters include back-propagation learning set (BP-set), learning rate (eta), momentum (alpha) and NN learning days (d_NN). Presented work is an extended version of previous work done by authors to integrate NN and similar days (SD) method for predicting day-ahead electricity prices. Similar days refer to similar price days, i.e., price data obtained from historical days. Euclidean norm with weighted factors is used for the selection of similar price days. Similar days method adopts the information of the days being similar to that of the forecast day. The similar days parameters, i.e., time framework of similar days (d = 45 days) and number of selected similar price days (N = 5) are kept constant for all the simulated cases. Forecasting performance is carried out by choosing a day from each season of the year 2006 randomly, and for which, the NN parameters for the base case are considered as BP-set = 500, eta = 0.8, alpha = 0.1 and d_NN = 45 days. Sensitivity analysis has been carried out by changing the value of BP-set (500, 1000, 1500); eta (0.6, 0.8, 1.0, 1.2), alpha (0.1, 0.2, 0.3) and d_NN (15, 30, 45 and 60 days). During the sensitivity analysis, the most favorable value of BP-set is first identified followed by that of eta and alpha, and based on which the best value of d_NN is determined. Sensitivity analysis results demonstrate that the best value of mean absolute percentage error (MAPE) is obtained when BP-set = 500, eta = 0.8, alpha = 0.1 and d_NN = 60 days for winter season, whereas for spring, summer and autumn, these values are 500, 0.6, 0.1 and 45 days, respectively. Results and discussions from real-world case study based on the PJM electricity market are presented.