Electrical load forecasting for commercial buildings using LSTM based algorithms

Short term load forecasting is essential in commercial buildings for sourcing power where price is time sensitive. Other algorithms developed for aggregated load at utility companies level, produce inaccurate results for specific bulk consumers like shopping malls, airports, universities etc. This paper presents a Long Short-Term Memory (LSTM) network combined with a fully connected layer architecture for short-term load forecasting (STLF). In this approach, after preprocessing the time series historical data, a new set of features are generated by lagging external weather variables and historical load values to capture temporal patterns. To fully assess our model’s performance in STLF, we conduct load forecasts for three short-term windows:one-hour-ahead,one-day-aheadandone-week-aheadforecasts. The performance of the proposed architecture is benchmarked against widely used models. The proposed model outperformed the baseline models by at least 7%, 9.37%, and 12.4% in MAE (Mean Absolute Error), MAPE (Mean Absolute Percentage Error), and RMSE (Root Mean Squared Error), respectively, for the one-week-ahead forecast. Similarly, it surpassed the baseline models by at least 9% in the one-day-ahead forecast and by at least 5% in the one-hour-ahead forecast across all metrics.