Techniques of applying wavelet de-noising into a combined model for short-term load forecasting

Abstract Electrical load time series are non-stationary and highly noisy because a variety of factors affect electrical markets. The direct forecasting of electrical load with noisy data is usually subject to large errors. This paper proposes a novel approach for short-term load forecasting (STLF) by applying wavelet de-noising in a combined model that is a hybrid of the seasonal autoregressive integrated moving average model (SARIMA) and neural networks. The process of the proposed approach first decomposes the historical data into an approximate part associated with low frequency and a detailed part associated with high frequencies via a wavelet transform. A SARIMA and a back propagation neural network (BPNN) are then established by the low-frequency signal to forecast the future value. Finally, the short-term load is forecasted by combining the prediction values of SARIMA and BPNN, and the weights of the combination are determined using a variance–covariance approach. To evaluate the performance of the proposed approach, the electricity load data in New South Wales, Australia, are used as an illustrative example. A comparison of the results with other models shows that the proposed model can effectively improve the forecasting accuracy.