Short Term Load Forecasting And Anomaly Detection in Residential Power Consumption

Load forecasting constitutes a fundamental component of the data analytics framework within the widely implemented advanced metering infrastructure. When compared with the aggregated load from a substation, forecasting the significantly fluctuating residential power consumption presents a formidable challenge. Various machine learning methodologies employed to address the general time series forecasting problem can be broadly categorized into statistical, probabilistic, hybrid, and deep learning models. Predominantly, load forecasting paradigms incorporate numerous extrinsic factors such as meteorological conditions, calendar events, and other temporally lagged attributes, which prove difficult to procure and archive at an individual site level. Our research undertakes a comparative analysis of contemporary transformer-based models and cutting-edge models like N-BEATS, N-HiTS, and TimesNet for the day-ahead prediction of the highly erratic residential power usage, drawing exclusively on historical load data, devoid of external features. Furthermore, given the impracticality of manual energy consumption monitoring, our study introduces a framework that amalgamates load forecasting with a rule-based anomaly detection mechanism. The outcomes of the regression models are systematically compiled, indicating that the TimesNet model surpasses other deep learning models in forecasting accuracy, notwithstanding its extensive training duration. This investigation was conducted utilizing publicly available authentic smart meter data from a residential setting.