Sensitivity learning oriented nonmonotonic multi reservoir echo state network for short-term load forecasting

Load forecasting is becoming an important issue day by day for economic generation of power, economic allocation between plants, maintenance scheduling and for system security which involves peak load shaving by power inter change with interconnecting utilities. In this paper, sensitivity learning oriented multi reservoir Echo State Network (ESN) using non monotonic transfer function with optimized structures by particle swarm optimization (PSO) algorithm, are used for short term load forecasting. Load time series of Electric Reliability Council of Texas (ERCOT) control area and Australian Energy Market Operator (AEMO) data are used for benchmarking the proposed method. Sensitivity oriented Linear Learning gives the sensitivities of the sum of squared errors. It has no extra computational cost, because the required information becomes available without having extra calculations. Echo state network parameters are being optimized with well-known Particle swarm optimization technique. Experimental results depicts that the proposed sensitivity oriented non monotonic Echo State network (SNESN) offers superior performance, in terms of mean absolute percentage error (MAPE), in time series prediction and eventually outperform the traditional load forecasting model like ARIMA and modern techniques like Support Vector Machine (SVM) based Genetic algorithm, Wavelet Neural Network and ANN based Fuzzy Network which prove the state of the art.

I. Introduction

Load forecasts with lead time varying from few hours to several years help in the efficient operation, control and planning of power systems. Load forecasting falls into three broad categories, which differ in time span as short-term, medium-term and long-term forecast [1]. Load prediction period may involve month or year for the long term and the medium term forecasts and day or may be hour for the short term forecast [2]. Because of the shrinking spinning reserve and the seasonal high-energy demand, utilities are dependent more than ever on short term forecasting for daily energy consumption [3]. Developments in the accuracy of short-term load forecasts can result in potential financial savings for utilities and co-generators. To achieve high forecasting accuracy and speed, which are the two most important requirements of Short-term load forecasting (STLF), it is important to analyse the load characteristics and identify the main factors affecting the load which will determine the generalization capability [4]. Traditional statistical load forecasting techniques, such as regression, time series, Fourier decomposition, have been used in practice for a long time. Though, these methods cannot significantly represent the complex nonlinear relationships that exist between the load and factors that influence it [5]. Modern load forecasting techniques, such as expert systems, Artificial Neural Networks (ANN), fuzzy logic, wavelets, have been developed recently, showing encouraging results [6]–[10]. Currently, dynamic and recurrent architectures constitute also an interesting research area. On the other hand Echo state network (ESN) has been drawing attention as a kind of artificial recurrent neural network (RNN). ESN facilitates the practical application of RNNs and has been proved that it can outperform classical fully trained RNNs in many tasks [11]–[14].