Load forecasting is an important action to achieve optimal power exchange and robust operation of utilities, especially in the active integration of distributed renewable energy sources (DRES). Deep learning methods have gained popularity among researchers due to their capability to generalize complex nonlinearity in power consumption fluctuations. Proper feature selection is one of the key factors in implementing a successful forecasting model. However, most of the works focus on temporal patterns of the data excluding spatial correlation of features. This work aims to select the features that have the highest impact on aggregate residential power usage and use these features to gain better forecasting accuracy. Manually extracted features from weather data provided significant changes in outputs. Evaluation of the RNN-based networks showed that they achieved better performance when input features were selected properly. MAE values of LSTM, Bidirectional LSTM, and GRU increased by 25, 30, and 11% respectively when the labeled day feature was added to the input date. Whereas the MAPE of the models was improved by 23, 30, and 13% respectively. The negative effect of the additional features proved that a robust technique must be applied in the feature extraction stage.