Accurate prediction of thermal load in buildings is essential for efficient energy planning. In this study, we investigate the application of Artificial Neural Networks (ANNs) to predict thermal load and indoor temperature evolution in residential buildings. We propose a flexible and adaptive model that is retrained on a daily basis to deliver updated hour-by-hour predictions for the following 24 h. To strike a balance between prediction accuracy and computational efficiency, we employ various design choices, including feature selection using Pearson's correlation coefficient (PCC), dynamic architecture, down-sampling, and specific state resetting and weight initialization. The results demonstrate the efficacy of our proposed model, as indicated by low root mean square error (RMSE) and mean bias error (MBE) values. For zone temperature, the average RMSE and MBE are 0.3 °C and 0.18 °C (summer) and 0.5 °C and 0.2 °C (winter), respectively. Furthermore, the average RMSE and MBE for thermal load predictions are 12 W/m2 and -2.4 W/m2 (winter) and 10 W/m2 and -0.6 W/m2 (summer), respectively. These performance metrics establish our model as a valuable tool for optimizing heating and cooling systems, resulting in energy savings and cost reductions. Our findings emphasize the potential of ANNs for precise thermal load and indoor temperature predictions, offering practical implications for building operators, engineers, and researchers involved in energy management.