Abstract: iming at the problems of insufficient utilization of multi-dimensional meteorological features and high computational overhead in multi-step time series modeling existing in short-term multi-step photovoltaic power forecasting, a multi-step photovoltaic power forecasting model integrating Informer encoder and SENet channel attention mechanism is proposed. Based on the Informer encoder, the model reduces the computational complexity of short-term multivariate time series modeling via probabilistic sparse self-attention. A SENet module is embedded at the front end of the network, which adaptively raises the feature weights of key meteorological factors such as irradiance through the squeeze-and-excitation recalibration mechanism. Meanwhile, multi-scale feature embedding, multi-dimensional time encoding and encoder information distillation strategies are introduced to strengthen the model’s capability of capturing intra-day multi-scale power fluctuations. Validations are carried out on measured photovoltaic power station datasets. The experimental results demonstrate that the proposed model outperforms comparative models including LSTM, Transformer and vanilla Informer in all evaluation metrics. It exhibits favorable forecasting robustness under various intra-day weather conditions such as sunny, cloudy and overcast days, and can satisfy the short-term multi-step forecasting requirements for intra-day dispatching of photovoltaic power stations.