Dynamic non-constraint ensemble model for probabilistic wind power and wind speed forecasting

Accurate and reliable probabilistic wind power and wind speed forecasts provide large amounts of uncertainty information, which is important for wind farm management and grid dispatch optimization. In this study, a dynamic non-constraint ensemble model is proposed to generate probabilistic wind power and wind speed forecasts. First, four deep Gaussian neural networks (DGNNs) based on popular time series forecasting approaches and the maximum likelihood estimation-based loss function are designed to generate base probabilistic forecasts in the ensemble model. Second, to consider the overall uncertainty of base probabilistic forecasts, a novel ensemble strategy for probabilistic forecasting is derived based on the probability density function of the weighted sum of finite Gaussian random variables. Third, to obtain the ensemble weights for different base probabilistic forecasts, a dynamic non-constraint weight learning model, containing quantile function, convolutional neural network, and channel attention, is proposed to generate dynamic non-constraint ensemble weights. In addition, the maximal information coefficient, which measures the linear and nonlinear relationship between the historical wind data and the target, is used for selecting the optimal input length. The experimental results from four real-world wind datasets demonstrate that the proposed ensemble model achieves exceptional accuracy in probabilistic wind power and wind speed forecasting. It outperforms DGNNs by an average improvement of 4.9325 % in pinball loss and surpasses Gaussian process regression by 16.6382 %. The effectiveness of utilizing non-constraint ensemble weights is supported by the results obtained with different weight constraints. Furthermore, hypothesis testing further confirms the overall effectiveness of the proposed ensemble model.