Forecasting the electricity load is crucial for power system planning and energy management. Since the season of the year, weather, weekdays, and holidays are the key aspects that have an effect on the load consumption, it is difficult to anticipate the future demands. Therefore, we proposed a weather-based short-term load forecasting framework in this paper. First, the missing data is filled, and data normalisation is performed in the pre-processing step. Normalization accelerates convergence and improves network training efficiency by preventing gradient explosion during the training phase. Then the weather, PV, and load features are extracted and fed into the proposed Highway Self-Attention Dilated Casual Convolutional Neural Network (HSAD-CNN) forecasting model. The dilated casual convolutions increase the receptive field without significantly raising computing costs. The multi-head self-attention mechanism (MHSA) gives importance to the most significant time steps for a more accurate forecast. The highway skip network (HS-Net) uses shortcut paths and skip connections to improve the information flow. This speed up the network convergence and prevents feature reuse, vanishing gradients, and negative learning problems. The performance of the HSAD-CNN forecasting technique is evaluated and compared to existing techniques under different day types and seasonal changes. The outcomes indicate that the HSAD-CNN forecasting model has low Mean Absolute Error (MAE), Mean Squared Error (MSE), Mean Absolute Percentage Error (MAPE), and a high R2.