Sequence-to-sequence neural networks for short-term electrical load forecasting in commercial office buildings

The U.S. power grid is transforming to become “smarter,” cleaner, and more efficient. This is leading to the addition of significant distributed variable renewable generation. Due to the variable nature of renewable generation, the short- and long-term supply-demand imbalances are less predictable, and conventional approaches to mitigating the imbalance will not be efficient or cost-effective. To address this challenge, transactive control technologies have been proposed. The transactive control approach requires individual end-use loads to express flexibility as a function of price. To model flexibility while maintaining robustness to any non-linear behavior exhibited by end-use loads, machine learning approaches for load forecasting are being explored. However, certain aspects, such as how much training data is required and how deep models for load forecasting should be structured and trained are not well understood. This work explores how to apply sequence-to-sequence recurrent neural networks to short-term electrical load forecasting with a case study of four commercial office buildings. We find that it is best to start the training in the middle of a heating or cooling season with at least six months of data. We further show that models perform best when predictions are conditioned on three to 12 h of prior data, with a decrease in performance for shorter contexts. We identify recommended ranges for common hyperparameters that could be used by practitioners applying similar models to their own tasks. Finally, we find that transferability of models across buildings is highly dependent on the building pairs, but in the best case, models are highly transferable.