Interpretable Load Forecasting with Structured State Space Neural Networks

Abstract

A central goal of load forecasting models is to produce reliable, trustworthy, and explainable predictions. Structured State Space Models have proven to be promising alternatives to recurrent neural networks such as the LSTM and transformers. However, most current load forecasting approaches, based on black-box neural network approaches, are hard to interpret and explain. We propose an efficient forecasting model based on stacked multi-input, multi-output state space sequence layers, enabling the analysis of the stability and system dynamics on a per-layer basis. The lightweight design reduces the number of parameters and operations by 67%, compared to its LSTM equivalent, and allows for deployment to low-cost edge devices with extremely tight resource constraints. The algorithms were tested with a one-year Smart Grid simulation and showed a decrease in RMSE by 7.8% compared to the LSTM equivalent.