Reliable solar power forecasting (SPF) is essential for effectively integrating and managing solar energy systems. However, conventional forecasting methods often encounter challenges when addressing the highly complex and nonlinear nature of solar power generation data. While recent advances in deep learning (DL), specifically convolutional neural networks (CNNs) and long short-term memory (LSTM) networks–offer improved forecasting performance, their transparency remains limited, undermining their real-world applications. This paper introduces a novel network that integrates CNNs, LSTMs, and attention mechanisms to enhance the accuracy and reliability of SPF. By integrating spatial feature extraction, temporal dependency capture, and focused attention on key time steps, we build a hybrid model (CNN-LSTM-Attention). Moreover, by feature ablation of explainable artificial intelligence (AI) solar power (XAI-SP) analyzing the model’s responses to different input features, we can pinpoint the most critical factors driving predictions and gain insight into its decision-making process. The experimental results demonstrate that the hybrid model achieves better performance compared to the baseline models on real-world datasets from two solar stations. Furthermore, feature ablation analysis (FAA) reveals that total solar irradiance and direct normal irradiance are critical features for achieving accurate forecasts. These findings can inform data acquisition strategies and feature selection for optimizing SPF models.