An improved encoder-decoder-based CNN model for probabilistic short-term load and PV forecasting

Abstract

Integrating distributed energy resources (DER) such as distributed generation, demand response, and plug-in electric vehicles is one of the major causes of fluctuating and unpredictable operating states in electric distribution systems. Therefore, distribution utilities must carry out anticipated operational planning to achieve appropriate and efficient network management. Then, it is necessary to obtain more accurate load forecasts on higher granularity levels than those commonly supervised by the SCADA system, for instance, at distribution transformers. Furthermore, as medium/low voltage profiles are more volatile and uncertain than high voltage profiles and, therefore, more difficult to predict, there is an opportunity to improve their performance at this level. This work proposes a short-term net load forecast model that considers load consumption and PV distributed generation behind the meter. This model is based on an efficient deep learning network that uses novel techniques and architectures implemented in other tasks adapted to the net electric load forecasting problem at an individual and/or low-aggregated level. At the same time, the model can consider information provided by exogenous variables of time and meteorological ones to improve the forecast accuracy. Additionally, the proposed model is extended to a probabilistic approach through Monte Carlo Dropout and kernel density estimation to obtain probability density forecasts. To evaluate the model performance, a dataset from the “Caucete Smart Grid” located in Argentina is used. The results show the effectiveness and superiority of the proposed model through several cases and comparisons with the state-of-the-art models considered.