Journal of Applied Research in Electrical Engineering

Journal of Applied Research in Electrical Engineering

Enhanced Electricity Demand Forecasting through Transparent Model Parameterization and Hyperparameter Optimization

Document Type : Research article

Authors
1 Department of Computer Science and Applications, SRM Institute of Science and Technology, Ramapuram Campus, Chennai, India
2 SRMIST, Chennai
Abstract
The power sector faces significant challenges due to the increase in electricity gadgets and usage. Gadgets and electricity appliances like air conditioning play an essential role all over the home and office sectors. For a few decades, the usage of electricity has increased. External factors like climatic conditions are also affecting the usage of electricity. Based on the economic growth of India, Tamil Nadu also increased in multi-growth industries. As of now, the usage of electric vehicles (two-wheelers as well as four-wheelers) is also rapidly increasing, which also leads to some fluctuation in electricity demand. These transformations change the patterns of electricity consumption and require careful management to ensure a stable supply. The suggested system uses XGBoost, a sophisticated gradient boosting technique suitable for tabular data and capable of handling nonlinearity. NeuralProphet handles time series-specific patterns (seasonality, trend) in a hybrid model, while XGBoost captures other patterns using designed features. This study introduces a new method for forecasting time series using NeuralProphet and XGBoost. NeuralProphet uses an additive model to model trends, seasonality, and event-based data. To improve Neural Prophet’s forecasts, residual errors generate more features, which are then fed into an XGBoost model for improved outcomes. XGBoost modeling then improves Neural Prophet’s ability to detect more nonlinear interactions that were not previously addressed. Finally, the predicting outputs from both models are combined, taking advantage of both NeuralProphet and FBProphet which is effective at catching reoccurrence and long-term oscillatory behaviors, and XGBoost, which explores the residual data structure.

Highlights

  • Proposed a hybrid NeuralProphet–XGBoost framework that combines transparent time-series decomposition with residual-based nonlinear learning for improved electricity demand forecasting accuracy.
  • Introduced residual-driven feature learning, where NeuralProphet forecast errors are explicitly modeled using XGBoost to capture complex nonlinear demand patterns.
  • Integrated weather, calendar, and socio-economic regressors within a unified forecasting pipeline to better reflect real-world electricity consumption behavior.
  • Provided transparent parameterization and hyperparameter tuning across all models, enabling reproducibility and clear interpretation of forecasting performance improvements.
  • Demonstrated consistent error reduction over standalone Prophet, FB Prophet, NeuralProphet, and XGBoost models on real-world electricity demand data.

Keywords
Subjects

Volume 4, Issue 2
Summer and Autumn 2025
July 2025
Pages 154-164

  • Receive Date 24 April 2025
  • Revise Date 14 December 2025
  • Accept Date 18 January 2026