Towards Continuous Power Forecasting Practical Continual Learning for Real-World Energy Systems in Nonstationary Time Series

Jul 16, 2026

Power forecasting models deployed in energy markets must handle nonstationary conditions where data distributions evolve with weather variability, infrastructure changes, and consumption behavior. The paper targets practical constraints including limited historical data for repeated retraining and the need for uninterrupted long-term service. It introduces Continuous Power Forecasting (CPF) as continual learning for regression, evaluates six representative approaches under realistic accessibility and update policies, and shows improved self-adaptation to drift while mitigating catastrophic forgetting without storing large histories.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages20

Reading Time50 min

Introduction

Related Work

Continuous Power Forecasting (CPF)

CLeaR Framework and Continual Learning Instantiations

Experimental Setup and Evaluation Protocol

Results Analysis and Conclusion

arXiv :2606 .24955v2 [ cs .LG] 10 Jul 2026 Towards Continuous Power Forecasting: Practical Continual Learning for Real-World Energy Systems in Nonstationary Time Series Yujiang He 1 (􀀀), Frederic Uhrweiller1 , and Bernhard Sick 1 Intelligent Embedded Systems, University of Kassel, 34121 Kassel, Germany {yujiang.he, frederic .uhrweiller, bsick}@uni-kassel .de Abstract. Power forecasting models deployed in real-world energy markets must operate under nonstationary conditions, where data distributions continually evolve due to weather variability, infrastructure upgrades, and changing consumption behaviors. In practice, these models face strict operational constraints: historical data may be limited for repeated retraining, and uninterrupted long-term service is often required. This paper addresses these challenges by proposing the paradigm of Continuous Power Forecasting, which views power forecasting as a continual learning problem rather than a static offline task. Based on an adaptive continual learning framework for regression, we systematically investigate the practical effectiveness of six representative continual learning approaches from three methodological categories. These approaches are evaluated under different realistic assumptions regarding data accessibility and update policies. Experimental validation on real-world power datasets demonstrates that continual learning enables forecasting models to self-adapt to distributional drift, accumulate knowledge over time, and mitigate catastrophic forgetting without relying on large-scale historical data storage. Beyond performance gains, our study provides practical insights into the stability and adaptation behaviors of different continual learning approaches under realistic operational constraints. Overall, this work illustrates how continual learning can be pragmatically integrated into industrial power forecasting pipelines, offering a scalable and sustainable solution for long-term deployment in dynamic environments. Keywords: Continual Learning · Power Forecasting · Time Series · Regression · Nonstationary. 1 Introduction Power forecasting is a core component of modern energy systems, supporting market operations and grid management. In real-world deployments, forecasting models must operate over long lifecycles under nonstationary conditions, where data distributions evolve due to changing weather patterns, infrastructure, and consumption behavior. At the same time, operational constraints such as limited access to historical data and the requirement for uninterrupted service make repeated offline retraining impractical. 2 Y. He et al. In practice, nonstationarity acts as both drifts in input distributions and changes in the underlying relationship between inputs and targets due to concept drift. While collecting large training datasets can partially mitigate such effects, this approach is costly, slow, and often incompatible with data privacy protection requirements. Conversely, naively fine-tuning models on incoming data leads to catastrophic forgetting, and this weakens long-term forecasting reliability. We address these challenges through the paradigm of Continuous Power Forecasting (CPF), which treats power forecasting as a lifelong adaptive process with the support of continual learning (CL) rather than a static offline task. Here, continuous emphasizes uninterrupted system-level operation over the model’s lifecycle, while continual denotes the repetitive learning mechanism that enables incremental adaptation and knowledge retention. The goal is to allow forecasting models to autonomously adapt to distributional drift and accumulate knowledge over time without relying on large-scale historical data storage. Based on the modular CLeaR (Continual Learning for Regression) framework [6], which enables unified comparison of distinct continual learning approaches for regression, this paper investigates practical effectiveness of six practica

FAQs

What benefits does continual learning provide for power forecasting in the paper?

Experimental results on real-world power datasets show models can self-adapt to distribution drift, accumulate knowledge over time, and mitigate catastrophic forgetting without relying on large-scale historical storage.

Jul 16, 2026

Upload by Ava Thompson