As renewable energy continues to penetrate modern power systems, accurate short-term load forecasting is crucial for optimizing power generation resource allocation and reducing operational costs. Traditional forecasting methods often overlook key factors such as holiday load variations and differences in user electricity consumption behavior, resulting in reduced accuracy. To address this, we propose an optimized short-term load forecasting method based on time and weather-fused features using a ConvLSTM-3D neural network. The Prophet algorithm is first employed to decompose historical electricity load data, extracting feature components related to time variables. Simultaneously, the SHAP algorithm filters weather variables to identify highly correlated weather features. A time attention mechanism is then applied to fuse these features based on their correlation weights, enhancing their impact within the time series. Finally, the ConvLSTM-3D model is trained on the fused features to generate short-term load forecasts. A case study using real-world data validates the proposed method, demonstrating significant improvements in forecasting accuracy.