Forecasting High Electricity Demand Using Explainable Machine Learning

The objective of this study was to predict the onset of high demand, the highest half-hourly average power, and the total duration of high demand over the next two hours using historical electricity consumption data from residential households. Half-hourly measurements from 2013 for 200 standard-tariff residential households selected from the Low Carbon London dataset were used. High demand was defined as exceeding the 90th percentile of each household’s average power values during the January–February period. The models were developed using data from 160 households; transfer performance was evaluated on the 40 households not included in model training. The process was divided into a training period (March–June), model selection in July, calibration of prediction intervals in August, and a testing period (September–December). The selected models were retrained using March–July data. Separate models were built for the three targets; Random Forest, histogram-based gradient boosting (HistGB), XGBoost, LightGBM, and CatBoost were compared against a weekly seasonal reference. Model selection was initially based on average precision (AP) for the target and mean absolute error (MAE) for power and duration. For the residential units used in the training, the HistGB model selected during the testing period produced an AP of 0.4183; the CatBoost models produced an MAE of 0.2704 kW for power and 9.2361 minutes for duration. The contributions of historical consumption and calendar features to the model predictions were examined using the SHAP method. In the early warning evaluation, 78.81% of relevant events in known residential units were detected in advance; 60.93% of early warnings were identified as false alarms. The improvement in duration prediction relative to a reference that yields a duration of zero in every window remained limited. The findings indicate that the approach could contribute to monitoring residential demand; however, the burden of false alarms and the limited success in duration prediction must be taken into account from a practical standpoint.

Authors

Institutions

Publication Details

Journal
Balkan Journal of Electrical and Computer Engineering
Published
2026-11-13
DOI
https://doi.org/10.17694/bajece.2060791
Primary Topic
Energy Load and Power Forecasting
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Forecasting High Electricity Demand Using Explainable Machine Learning

Süleyman Dal
Balkan Journal of Electrical and Computer Engineering
Energy Load and Power Forecasting
article

Forecasting High Electricity Demand Using Explainable Machine Learning

Süleyman Dal
article en

Abstract

The objective of this study was to predict the onset of high demand, the highest half-hourly average power, and the total duration of high demand over the next two hours using historical electricity consumption data from residential households. Half-hourly measurements from 2013 for 200 standard-tariff residential households selected from the Low Carbon London dataset were used. High demand was defined as exceeding the 90th percentile of each household’s average power values during the January–February period. The models were developed using data from 160 households; transfer performance was evaluated on the 40 households not included in model training. The process was divided into a training period (March–June), model selection in July, calibration of prediction intervals in August, and a testing period (September–December). The selected models were retrained using March–July data. Separate models were built for the three targets; Random Forest, histogram-based gradient boosting (HistGB), XGBoost, LightGBM, and CatBoost were compared against a weekly seasonal reference. Model selection was initially based on average precision (AP) for the target and mean absolute error (MAE) for power and duration. For the residential units used in the training, the HistGB model selected during the testing period produced an AP of 0.4183; the CatBoost models produced an MAE of 0.2704 kW for power and 9.2361 minutes for duration. The contributions of historical consumption and calendar features to the model predictions were examined using the SHAP method. In the early warning evaluation, 78.81% of relevant events in known residential units were detected in advance; 60.93% of early warnings were identified as false alarms. The improvement in duration prediction relative to a reference that yields a duration of zero in every window remained limited. The findings indicate that the approach could contribute to monitoring residential demand; however, the burden of false alarms and the limited success in duration prediction must be taken into account from a practical standpoint.

Balkan Journal of Electrical and Computer EngineeringVol. 14
Batman University (TR)
Affordable and clean energy, Climate action
Energy Load and Power Forecasting
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon