A deployable electrical load forecasting solution for commercial buildings

Electrical load forecasting for buildings plays a key role in the smart-grid paradigm, since accurate forecasts help in efficient energy management. This paper presents a practical solution, which can be readily deployed, by considering various real-life issues. Load forecasting algorithms are developed using Non-linear AutoRegressive model with eXogenous input (NARX) neural network and Support Vector Regression (SVR) to forecast the power consumption for day ahead, week ahead and month ahead at 15 minute granularity. Missing values, outliers in the power consumption data are treated using simple but effective techniques based on the thorough understanding of the power consumption time-series data. Autoregressive features, temperature and some efficacious contextual information, which are completely pertinent to this problem, are being derived to model this heterogeneous time series that consists of disparate patterns on weekdays, weekends and holidays. The novelty of the solution lies in the fact that it extensively models all the variations and produces accurate long term predictions at high granularity (15 min). The algorithms are validated using the actual power consumption data from three Office buildings; forecasting accuracy of both SVR and NARX neural network are comparable, the latter performs better by a thin margin. The average forecasting accuracy is nearly 93%, 88-90% and 85-87% for day ahead, week ahead and month ahead forecasting respectively.