Multi-step photovoltaic power forecasting using TCN and enhanced feature extraction with TimesNet

Abstract The growing integraion of photovoltaic (PV) systems in buildings poses significant challenges for energy management due to the intrinsic volatility and intermittency of PV output. Accurate multi-step PV forecasting is critical for supporting building energy management systems in scheduling, load optimization, and renewable energy dispatch. To tackle the dual challenges of modeling complex temporal dynamics and improving multi-step forecasting accuracy, this study introduces a channel-attention-based hybrid prediction framework, TCN-iTimesNet. The framework first leverages a temporal convolutional network (TCN) to capture long-range dependencies and extract stable global representations, which are subsequently processed by an improved TimesNet module for multi-scale periodic modeling. A convolutional kernel decoupling mechanism disentangles intra- and inter-period convolutions, enabling precise extraction of fine-grained features and long-range coupling patterns. In parallel, integration of a squeeze-and-excitation channel attention mechanism adaptively amplifies critical features, boosting both forecasting accuracy and model robustness. Experiments on a publicly available photovoltaic dataset from China indicate that the proposed model consistently outperforms mainstream baselines—including Transformer, TimesNet, CNN-LSTM, Informer, and DLinear—in predictive accuracy and long-horizon forecasting performance. Cross-dataset validation further demonstrates that TCN-iTimesNet maintains strong robustness and generalization across different PV datasets. The results validate the effectiveness of the proposed framework in multi-step photovoltaic power forecasting, providing a robust data foundation for building energy management and renewable energy dispatch planning.