A photovoltaic power forecasting framework based on Attention mechanism and parallel prediction architecture

Photovoltaic power generation is susceptible to the stochastic volatility characteristics of meteorological conditions, so it is of great significance to forecast the photovoltaic power generation accurately and reliably. This paper proposes a novel hybrid forecasting framework (Attention-DCC-BiLSTM-AR model, ADBA model) for ultra-short-term photovoltaic power prediction, which combines the Attention mechanism and a well-designed parallel prediction architecture with linear Autoregressive (AR) component and nonlinear Dilated Causal Convolution-Bidirectional Long Short-Term Memory network (DCC-BiLSTM) component. Firstly, Attention mechanism is employed to assign weights to input variables according to their relative importance, so as to optimize the multivariate time series. Secondly, the optimized data is fed into linear and nonlinear components of the parallel architecture for prediction, respectively. The nonlinear prediction component is implemented by a combined DCC-BiLSTM structure, which has complementary strength in extracting spatial and temporal features. Subsequently, the extracted features are fed into feature mapping layers to obtain the nonlinear fitting results. The linear prediction component is implemented by a statistical AR model, which can mitigate the scale sensitivity problem associated with neural networks and provide the linear fitting results. This parallel prediction architecture enables the hybrid framework to model both linear and nonlinear characteristics of historical power generation time series simultaneously. Finally, the prediction results of two components are integrated to obtain the final prediction result. Experimental results demonstrate that: the proposed model consistently outperforms benchmark models in terms of forecasting accuracy and robustness, and has shown the most superior prediction performance on different sites, different seasons, and different prediction time horizons.