PSCNet: Long sequence time-series forecasting for photovoltaic power via period selection and cross-variable attention

With the continuous expansion of photovoltaic installation capacity, accurate prediction of photovoltaic power generation is crucial for balancing electricity supply and demand, optimizing energy storage systems, and improving energy efficiency. With the help of deep learning technologies, the stability and reliability of the photovoltaic power generation prediction have been significantly improved. However, existing methods primarily focus on temporal dependencies and often fall short in capturing the multivariate correlations between variables. In this paper, we propose a novel long-sequence time-series forecasting network for photovoltaic power via period selection and Cross-variable attention, named PSCNet. Specifically, we first propose the Top-K periodicity selection module (TPSM) to identify the Top-K principal periods for decoupling overlapped multi-periodic patterns, enabling the model to attend to periodic changes across different scales simultaneously. Then, we design a time-variate cascade perceptron to capture both temporal change patterns and variate change patterns in the time series. It contains two elaborate modules named Time-mixing MLP (TM-MLP) and Cross-variable Attention Module (CvAM). The former module aims to capture long-short term variations in time series while the latter one integrates the effective information from different auxiliary variates that have an impact on photovoltaic power forecasting to enhance the feature representation for better power prediction. Extensive experiments on the DKASC, Alice Springs dataset demonstrate that our model can outperform existing state-of-the-art photovoltaic power forecasting methods in terms of three common-used metrics including Mean Average Error (MAE), Mean Squared Error (MSE), and Mean Absolute Percentage Error (MAPE).