HOUSE HOLD ELECTRICITY CONSUMPTION PREDICTION USING ML

Abstract

Household electricity consumption has grown steadily with the wider use of air conditioners, refrigerators, water heaters, washing machines, and electronic devices. Most households learn how much energy they have used only when the monthly bill arrives, by which time it is too late to change their behaviour for that billing period. Distribution companies also need reliable estimates of future demand at the level of feeders and transformers in order to plan supply and avoid overloading. This paper presents a household electricity consumption prediction system that uses machine learning to forecast energy use over the coming hours, days, and billing period. The system takes its data from two sources. The first is a public household power consumption dataset recorded at one-minute resolution, containing active power, reactive power, voltage, current, and sub-metering readings for different groups of appliances. The second is a small smart-metering setup built by the project team using an ESP32 microcontroller and a non-invasive current transformer sensor with a voltage sensing module, which measures the load of a real household and sends readings over Wi-Fi to a server. Weather data such as outdoor temperature and humidity and calendar information such as weekends and holidays are added because they strongly affect domestic load. The raw readings are cleaned to remove missing intervals and measurement errors, aggregated to hourly and daily totals, and converted into features describing recent usage, time of day, day of week, season, and temperature. Several regression models are trained and compared, including Linear Regression, Decision Tree, Random Forest, Gradient Boosting, XGBoost, and a Long Short-Term Memory network. The models are evaluated with a time-ordered split, and performance is measured using RMSE, MAE, MAPE, and the coefficient of determination. The results show that gradient boosting and LSTM models capture daily and weekly patterns well, with lagged consumption and outdoor temperature as the most important inputs. The predictions are presented on a web dashboard that shows real-time power, daily consumption, the forecast for the next day, and an estimate of the monthly bill based on the applicable tariff slabs. The dashboard also displays how consumption is divided among appliance groups and highlights unusual usage, such as a sudden rise in night-time load, which may indicate a faulty appliance or equipment left on unintentionally. The system is useful for households that wish to manage their electricity use and budget, and it also illustrates how aggregated forecasts can support demand planning by distribution companies. Future work includes appliance-level load disaggregation, integration with rooftop solar generation data, and demand-response features that suggest shifting flexible loads to periods of lower tariff.