Long-Term Electricity Demand Prediction Using Non-negative Tensor Factorization and Genetic Algorithm-Driven Temporal Modeling

TL;DR

This paper tackles long-term electricity demand forecasting using only historical consumption data, avoiding external covariates. It integrates Non-negative Tensor Factorization (NTF) on a third-order tensor $\mathcal{X} \in \mathbb{R}^{I\times J\times K}$ with a Genetic Algorithm (GA) that optimizes ARIMA hyperparameters for each latent annual factor, yielding $\hat{\mathcal{X}} = \sum_{r=1}^R a_r \circ b_r \circ \hat{c}_r$. Empirical results on Japanese data show that the NTF+GA framework achieves lower MSE than baselines that skip NTF or GA, and that enforcing latent structure reduces overfitting while enabling interpretable temporal patterns. The study also reveals that initialization randomness in NTF can affect reconstructions, though robustness is attainable through multiple initializations or ensemble strategies. Overall, the approach is scalable, covariate-free, and extendable to other structured time-series forecasting tasks with tensor representations.

Abstract

This study proposes a novel framework for long-term electricity demand prediction based solely on historical consumption data, without relying on external variables such as temperature or economic indicators. The method combines Non-negative Tensor Factorization (NTF) to extract low-dimensional temporal features from multi-way electricity usage data, with a Genetic Algorithm that optimizes the hyperparameters of time series models applied to the latent annual factors. We model the dataset as a third-order tensor spanning electric utilities, industrial sectors, and years, and apply canonical polyadic decomposition under non-negativity constraints. The annual component is forecasted using autoregressive models, with hyperparameter tuning guided by the prediction error or reconstruction accuracy on a validation set. Comparative experiments using real-world electricity data from Japan demonstrate that the proposed method achieves lower mean squared error than baseline approaches without tensor decomposition or evolutionary optimization. Moreover, we find that reducing the model's degrees of freedom via tensor decomposition improves generalization performance, and that initialization sensitivity in NTF can be mitigated through multiple runs or ensemble strategies. These findings suggest that the proposed framework offers an interpretable, flexible, and scalable approach to long-term electricity demand prediction and can be extended to other structured time series forecasting tasks.

Figures (9)

• Figure 1: Relationship between CP decomposition rank and reconstruction accuracy.
• Figure 2: Prediction results for Pattern A2010.
• Figure 3: Prediction results for Pattern B2010.
• Figure 4: Prediction results for Pattern A2000.
• Figure 5: Prediction results for Pattern B2000.