Adaptive data predictions for the energy sector at national level

Electric load forecasting is a central aspect of power generation planning as it allows the optimization of the production units. To date, no artificial neural network architectures (ANNs) were found that can precisely predict the consumption of energy at national level. In this paper, we propose the implementation of an artificial intelligence forecasting model that focuses on short term predictions. The current version is an enhancement of the previous approach which consisted of a full implementation in MATLAB. The algorithm was transposed in Python, using the new and updated tools such as TensorFlow and Keras, while taking into consideration a performance comparison between the two. To validate our model, we used, data from Romania between years 2008 to 2011. The implementation focuses on four main stages: restructuring and pre-processing the data, finding, and training the optimal model, refining the initial model, and retraining the neural network with new data. In terms of results, the current implementation decreased considerably the training time and returned a good prediction capability. On the other hand, the Python model was prone to overfitting, problem that was solved with techniques such as dropout and regularization layers. Regarding the architecture, it uses classical neurons as compared to other approaches in time series prediction that use LSTM cells. This simpler neural network offered higher efficiency in terms of computational resources while also being able to make accurate predictions.