Ensembling Approaches to Hierarchical Electric Load Forecasting

Short term electrical load forecasting is critical in ensuring reliability and operational efficiency for electrical systems. With an influx of monitoring data and the growing technical complexity of the grid, there is a great interest and need for accurate forecasting in electricity planning. Our project uses a curated electric load dataset from Kaggle and evaluates the performance of several different load forecasting methods. We first implement a simple parametric regression, where we divide the problem into subproblems by key indicator variables and fit multiple regressions. Second, we use a similar weather load input similar to Chen et. al [3] and fit a Neural Network. Third, we decompose the load into low and high frequency profiles and fit two separate networks to predict these components. Last, we evaluate linear combinations of these models to optimize performance on the validation set.