Short-term wind speed forecasting framework based on stacked denoising auto-encoders with rough ANN

Abstract

In this paper, a multi-modal short-term wind speed prediction framework has been proposed based on Artificial Neural Networks (ANNs). Given the stochastic behavior and high uncertainty of wind speed, in modern power systems with high penetration of wind power, precise wind speed forecasting is necessary for the power utility managers. To ensure that the prediction results are sufficiently precise, in this study, a multi-modal method is designed based on denoising and prediction modules. In denoising module, at first, the optimal configuration of Stacked Denoising Auto-encoders (SDAEs) is determined based on Genetic Algorithm (GA), and, then, SDAEs are applied for pre-training the data to reduce the input data noise. Due to the fact that different data sources have different noise bases, the optimal structure of the denoising module is arranged completely independent of various input data. In the forecasting module, the Sinusoidal Rough-Neural Network (SR-NN) is utilized to predict wind speed. After that, denoising and forecasting modules are stacked together to make the holistic deep structure network. In this study, the Deep Learning (DL) approach has been utilized to extract more robust features from the input data in the denoising process. To handle the high intermittent behavior of wind speed, the rough neurons are considered. Rough neurons are made up of two pairs of conventional neurons that are known to be upper and lower bound neurons. Since the wind speed has a periodic and nonlinear behavior, an ANN, based on the sinusoidal activation function, is more accurate than the typical type of sigmoid activation function. In this paper, weather data from Ahar, Iran, which has a high potential wind power, is considered. In modern power systems, the extra short-term forecasting is also needed; therefore, in this study, wind speed forecasting in one hour and 10-minute intervals are employed. Input data are selected based on Grey Correlation Analysis (GCA), and according to the GCA result, in addition to wind speed data at various heights, environmental data such as humidity and temperature are also considered. To evaluate the efficiency of the proposed method, the simulation results are compared with other structures of ANNs and benchmarking machine learning methods such as Support Vector Machine (SVM) and Autoregressive Moving Average (ARMA) methods in different scenarios. In this comparison, the impact of DL has been thoroughly investigated. In this paper, the comparison of the performance of deep and shallow network structures has been studied.