Interval-based Probabilistic Load Forecasting for Individual Households: Clustering Approach

With the widespread use of smart meters, it has become easier to manage demand side at the individual house-hold level by employing applications such as load forecasting. However, uncertainty in the load consumption profiles is a major challenge for individual load forecasting methods caused by the key factors such as variation in user behavior, and weather variables. Therefore, the load profiles first need to be modeled systematically in order to achieve effective forecasting results. This paper presents a holistic load forecasting framework by first modeling the temporal features of load consumption profiles using Gaussian mixture model clustering. The extracted information is then fed to the Bayesian Bidirectional long short-Term memory (LSTM) method to generate probabilistic forecasts. The proposed framework is implemented on real-life energy consumption data and compared against benchmark machine learning methods using forecasting evaluation metrics at 90%, 50%, and 10% quantiles.