Interpretable Net Load Forecasting Using Smooth Multiperiodic Features

Manuscript, posted April 2024.

• Current version

• Code

We consider the problem of forecasting net load over a future horizon such as one day, using a trailing window of past net load values as well as date and time. We focus on three variations on this problem: point forecasts, marginal quantile forecasts, and generating conditional samples of the future values. These tasks can be accomplished using methods that range from basic, such as linear regression models, to sophisticated ones involving trees or neural networks. We propose a method that relies on linear regression using some custom engineered time-based features to capture multiple periodicities, such as daily, weekly, and seasonal, and their interactions, such as the variation in daily patterns over the year. Our proposed models are readily interpretable, and rely on efficient and reliable convex optimization to fit. At the same time, the method has strong predictive power, outperforming baseline techniques, and gracefully supports missing data. We illustrate our method on three years of hourly net load data for the state of Rhode Island, comparing predictions made with various subsets of the features. We provide an open source implementation that can be used for any time series that exhibits multiple periodicities.