EXPLAINABLE DEEP LEARNING ENERGY DEMAND FORECASTING SYSTEMS

EMMANUEL, CHUKWUDIEBUBE ONYENSO (2026) EXPLAINABLE DEEP LEARNING ENERGY DEMAND FORECASTING SYSTEMS. Other thesis, GODFREY OKOYE UNIVERSITY, ENUGU,.

Text EMMANUEL_CHUKWUDIEBUBE_ONYENSO_MAIN_WORK-1.pdf Download (1MB)

Abstract

Modern power systems face intense pressures to consistently supply electricity against the backdrop of increasingly complicated electricity demand and thus precise electricity consumption forecasting is an ever more important issue. Traditional statistical methods (notably ARIMA, linear regression) have consistently had limitations in capturing non-linear, time-dependent patterns inherent in hourly electricity demand data, resulting in forecasts that fail to achieve an adequate level of accuracy for effective use in real-time grid operation. Deep learning models developed for electricity forecasting have largely acted as "black box" models, producing forecasts with no justification of the variables which influence the forecast. This work is to overcome these two issues by the design and implementation of an Explainable Deep Learning Architecture for Hourly Energy Demand Forecasting which is based on a stacked Long Short-Term Memory neural network together with post-hoc SHAP (SHapley Additive exPlanations) explainability technique. This model was trained and tested using data obtained from a real US energy system (PJM Interconnection) electricity consumption hourly dataset consisting of over 121,000 data points. An end-to-end system was developed which involved preprocessing, temporal feature extraction, sequence creation, LSTM model training with dropouts for regulation, and SHAP based analysis for model explainability. The trained LSTM model resulted in a Mean Absolute Error (MAE) of 168.28 MW, a Root Mean Squared Error (RMSE) of 220.70 MW, and a coefficient of determination (R2) of 0.9921 on the test data. These were well beyond the predictive capabilities of all other competing methods. Analysis via SHAP indicated that the immediately previous hour's consumption (t-1_lag1) was by far the most significant predictor for current hour's consumption with the next shortest-range lags and the 24-hour lagged value also found to be significant predictors. Thus, it has been demonstrated that integrating deep learning with model explainability approaches can be used to produce a technically competent as well as credible predictive system for smart grids.

Item Type:	Thesis (Other)
Subjects:	Q Science > Q Science (General)
Divisions:	Faculty of Computing And Information Technology (FACIT)
Depositing User:	Nnenna Ayo
Date Deposited:	27 Jul 2026 09:10
Last Modified:	27 Jul 2026 09:10
URI:	http://eprints.gouni.edu.ng/id/eprint/6021

Actions (login required)

View Item