Modeling and clustering of residential electricity consumption for short-term forecasting using seasonal autoregressive integrated moving average (SARIMA)

Power demand is increasing in all sectors of the developing nations; however, the inadequate and disproportionate generation and distribution resources create a burden for the authorities to efficiently manage the power grid. This results in network overloading and consequent blackouts. To efficiently manage the resources, short-term consumption modeling (using physical or data-based models) is mandatory. Physical models, in contrast to the data-based models, are generally complex and difficult to be used for automatic control strategies for the grid. However, for the latter, there is generally a lack of data availability in the developing nations, as energy measurements are taken just once a month for billing purposes. A recent study conducted consumption measurements, at a high sampling rate, of 42 residential houses from different neighborhoods in Lahore, Pakistan. As mentioned earlier, such a dataset is rare in the developing nations, therefore, leverage the availability of this data set and perform SARIMA (seasonal autoregressive integrated moving average) and ARIMA (autoregressive integrated moving average) based time series modelling of the consumption. Furthermore, for a wider applicability of the results, this work also includes clustering of the houses based on the parameters such as power consumption in different seasons, number of occupants, structural attributes, demographics, and household appliances. Hierarchical clustering with complete linkage was used to derive household groups, Euclidean distance was applied as the similarity metric, and silhouette analysis validated cluster compactness and separation. We report models of different groups of residential houses, categorized with respect to the aforementioned parameters, to facilitate the operators for efficient power management. The findings in this paper are supported by a variety of simulation results, across 41 households, the dataset was split into training and testing sets, corresponding to 93%–7% for the 24-hour horizon and 82%–18% for the 72-hour horizon, which showed that SARIMA consistently outperformed ARIMA, reducing errors for 24-h (MAE: 0.30→0.23 kW, RMSE: 0.40→0.32 kW, MAPE: 36.9%→28.3%) and 72-h horizons (MAE: 0.39→0.32 kW, RMSE: 0.51→0.41 kW, MAPE: 49.1%→39.54%) and yielding lower AICc values. Clustering validation using silhouette widths (0.13–0.25) confirmed the robustness of identified household groups.