Very short-term spatial and temporal wind power forecasting: A deep learning approach

Abstract

In power systems that experience high penetration of wind power generation, very short-term wind power forecast is an important prerequisite for look-ahead power dispatch. Conventional univariate wind power forecasting methods at presentonly utilize individual wind farm historical data. However, studies have shown that forecasting accuracy canbe improved by exploring both spatial and temporal correlations among adjacent wind farms. Current research on spatial-temporal wind power forecasting is based on relatively shallow time series models that, to date, have demonstrated unsatisfactory performance. In this paper, a convolution operation is used to capture the spatial and temporal correlations among multiple wind farms. A novel convolution-based spatial-temporal wind power predictor (CSTWPP) is developed. Due to CSTWPP's high nonlinearity and deep architecture, wind power variation features and regularities included in the historical data can be more effectively extracted. Furthermore, the online training of CSTWPP enables incremental learning, which makes CSTWPP non-stationary and in conformity with real scenarios. Graphics processing units (GPU) is used to speed up the training process, validating the developed CSTWPP for real-time application. Case studies on 28 adjacent wind farms are conducted to show that the proposed model can achieve superior performance on 5-30 minutes ahead wind power forecasts.