Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cyclical Temporal Encoding and Hybrid Deep Ensembles for Multistep Energy Forecasting

Salim Khazem, Houssam Kanso

TL;DR

This work tackles multistep short-term electricity load forecasting by integrating cyclical temporal encodings with a hybrid LSTM-CNN ensemble and horizon-specific meta-learners. It introduces a systematic time-feature engineering pipeline with sine/cosine transforms and correlation-based feature selection, evaluated on a 12-month French national dataset. The results show consistent improvements across seven forecast horizons, with the hybrid ensemble outperforming single architectures and baselines. The study advances interpretable temporal representations in energy forecasting and demonstrates practical gains for demand management in smart grids.

Abstract

Accurate electricity consumption forecasting is essential for demand management and smart grid operations. This paper introduces a unified deep learning framework that integrates cyclical temporal encoding with hybrid LSTM-CNN architectures to enhance multistep energy forecasting. We systematically transform calendar-based attributes using sine cosine encodings to preserve periodic structure and evaluate their predictive relevance through correlation analysis. To exploit both long-term seasonal effects and short-term local patterns, we employ an ensemble model composed of an LSTM, a CNN, and a meta-learner of MLP regressors specialized for each forecast horizon. Using a one year national consumption dataset, we conduct an extensive experimental study including ablation analyses with and without cyclical encodings and calendar features and comparisons with established baselines from the literature. Results demonstrate consistent improvements across all seven forecast horizons, with our hybrid model achieving lower RMSE and MAE than individual architectures and prior methods. These findings confirm the benefit of combining cyclical temporal representations with complementary deep learning structures. To our knowledge, this is the first work to jointly evaluate temporal encodings, calendar-based features, and hybrid ensemble architectures within a unified short-term energy forecasting framework.

Cyclical Temporal Encoding and Hybrid Deep Ensembles for Multistep Energy Forecasting

TL;DR

This work tackles multistep short-term electricity load forecasting by integrating cyclical temporal encodings with a hybrid LSTM-CNN ensemble and horizon-specific meta-learners. It introduces a systematic time-feature engineering pipeline with sine/cosine transforms and correlation-based feature selection, evaluated on a 12-month French national dataset. The results show consistent improvements across seven forecast horizons, with the hybrid ensemble outperforming single architectures and baselines. The study advances interpretable temporal representations in energy forecasting and demonstrates practical gains for demand management in smart grids.

Abstract

Accurate electricity consumption forecasting is essential for demand management and smart grid operations. This paper introduces a unified deep learning framework that integrates cyclical temporal encoding with hybrid LSTM-CNN architectures to enhance multistep energy forecasting. We systematically transform calendar-based attributes using sine cosine encodings to preserve periodic structure and evaluate their predictive relevance through correlation analysis. To exploit both long-term seasonal effects and short-term local patterns, we employ an ensemble model composed of an LSTM, a CNN, and a meta-learner of MLP regressors specialized for each forecast horizon. Using a one year national consumption dataset, we conduct an extensive experimental study including ablation analyses with and without cyclical encodings and calendar features and comparisons with established baselines from the literature. Results demonstrate consistent improvements across all seven forecast horizons, with our hybrid model achieving lower RMSE and MAE than individual architectures and prior methods. These findings confirm the benefit of combining cyclical temporal representations with complementary deep learning structures. To our knowledge, this is the first work to jointly evaluate temporal encodings, calendar-based features, and hybrid ensemble architectures within a unified short-term energy forecasting framework.

Paper Structure

This paper contains 16 sections, 3 equations, 5 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Preprocessing & Feature Engineering
  5. Exploration of Time data granularity
  6. Feature Engineering for Cyclical Pattern of Time Series
  7. Experiments
  8. Principles
  9. Implementation Details and Dataset
  10. Base Classifier Structures and Configuration
  11. Results
  12. Impact of Time-Based Feature Engineering
  13. Forecasting Performance
  14. Ablation Study
  15. Comparison with Previous Literature
  16. ...and 1 more sections

Figures (5)

  • Figure 1: Power consumption variation along the year: (a) grouped by month; (b) grouped by week; (c) grouped by day in week; (d) grouped by day type Note: In this figure, “day type” refers exclusively to business day vs. weekend.
  • Figure 2: Comparison between raw and transformed temporal features. (Top) Normalized power consumption plotted with the raw Week in Year feature, showing discontinuities at year boundaries. (Bottom) Same power series plotted with its cyclical encodings: Week-Cos and Week-Sin, which produce smooth and continuous representations across years.
  • Figure 3: Workflow of the Ensemble-Learning Method for Power Forecasting
  • Figure 4: Meta Classifier based on Multiple MLPRegressors
  • Figure 5: Error metrics across the 7-day forecast horizon: (a) Mean Absolute Error (MAE), (b) Root Mean Square Error (RMSE).