CNN-BiLSTM-Attention Model for Wind and Solar Power Forecasting and Analysis of Factors Affecting Forecast Accuracy

With the annual increase in renewable energy installed capacity, accurate wind and solar power forecast models are essential to support the operation of power systems. To achieve precise forecasts, this paper proposes an ultra-short-term power forecast method based on CNN-BiLSTM-Attention. Firstly, hub wind speed and direct solar irradiance are selected as typical meteorological factors for wind and solar power forecast. Secondly, a forecast model is constructed by integrating convolutional neural networks (CNN) and bidirectional long short-term memory networks (BiLSTM), along with an attention mechanism. Comparative experimental results show that the proposed model outperforms traditional single forecast models across three evaluation metrics. Additionally, the paper analyzes the impact of data cleaning and data volume on forecast accuracy, finding that data cleaning effectively removes outliers, improving forecast accuracy by up to 16%. Therefore, with limited computational resources, priority should be given to enhancing the quality of the model training database, avoiding the use of large amounts of redundant data for model training.