Prediction Intervals with LSTM Networks Trained by Joint Supervision

This paper presents an approach for prediction interval generation by training a LSTM neural network with a joint supervision Loss Function. The prediction interval model provides the expected value and the upper and lower bounds of the interval given a desired coverage probability. The prediction interval models based on LSTM networks are compared with the classical recurrent neural network approach and are tested using two case studies. The first case corresponds to the forecasting up to one day ahead of the demand profile of 20 dwellings from a town in the UK, and the second case corresponds to the net power from an energy community made up 30 dwellings with a 50% level of photovoltaic power penetration. By using LSTM networks as the backbone of the proposed architecture, high-quality intervals are obtained with a narrower interval width compared with the classical recurrent neural network approach. Furthermore, the information provided by the prediction interval based on the LSTM network could be used to develop robust energy management systems that, for example, consider the worst-case scenario.