PV-MLP: A lightweight patch-based multi-layer perceptron network with time–frequency domain fusion for accurate long-sequence photovoltaic power forecasting

With the rapid development of the photovoltaic (PV) industry, achieving accurate long-sequence PV power forecasting is important for improving grid operation efficiency. However, as the number of time steps increases, the prediction errors also accumulate, making accurate long-sequence forecasting a significant challenge. Moreover, although the mainstream forecasting models based on the Transformer architecture can achieve satisfactory forecasting accuracy in long-sequence forecasting, they often encounter significant computational burdens, making them difficult in practical engineering deployments. To this end, based on the patch technique and multi-layer perceptron (MLP) structure, we propose two lightweight forecasting models: PV-MLP and PV-MLPx. Specifically, the proposed methods first correct data anomalies and reduce feature redundancy. Then, in the proposed learning models, we first segment the time-series data into multiple patches. PV-MLP employs a shared MLP layer to extract the temporal features of each patch, which effectively reduces the computational burden. PV-MLPx processes each patch independently using dedicated MLP layers, making it suitable for scenarios with abundant computational resources. Finally, the features extracted from each patch are fused along the feature dimension and generate the final results directly through a fully connected layer. Additionally, to address the issue of ignoring label sequence autocorrelation in the direct prediction paradigm, we build upon the traditional mean squared error loss function by applying the Fast Fourier Transform (FFT) to transfer the predicted and ground truth values to the frequency domain for secondary alignment. Finally, we conducted a comprehensive evaluation of PV-MLP’s forecasting accuracy and generalizability. Experimental results show that our proposed model achieves over 68.7%, 43.1%, and 38.4% improvements in Mean Squared Error (MSE) compared to the baseline models in data 1 on three long-term forecasting tasks with horizons of 96, 192, and 384 steps, respectively. Moreover, the Mean Absolute Error (MAE) of our model remains below 1.0 on three forecasting tasks, significantly outperforming the other baseline models.