Accurate electricity load forecasting is critical for efficient energy management, yet the interplay of weather variables and model interpretability remains underexplored. This study presents a CNN-based framework for short-term load forecasting, integrating hyperparameter optimization and SHapley Additive exPlanations (SHAP) for model transparency. Using historical weather data (temperature, humidity, precipitation) and load records, we first preprocess the data into temporally coherent sequences and normalize features to address scale variance. A tunable CNN architecture, optimized through Hyperband, achieves a validation loss of 1.146 MW, demonstrating robust predictive performance. SHAP analysis is then applied to quantify feature contributions, revealing temperature as the dominant predictor and uncovering nonlinear interactions between humidity and load demand. The model achieves an RMSE of 12.4 MW and R2 of 0.93 on test data, outperforming baseline methods. By combining predictive accuracy with interpretability, this work advances trustworthy AI for energy systems, enabling utilities to prioritize critical weather factors and refine grid strategies.