A survey and benchmark for household electricity forecasting: From statistical to foundation models

Energy forecasting plays a crucial role in modern power systems, enabling accurate estimation of energy consumption and generation to support cost reduction, grid stability, and the integration of distributed energy resources such as photovoltaic systems, electric vehicles, and battery energy storage systems (BESS). While forecasting methodologies have evolved from traditional statistical and machine learning models to advanced transformer-based architectures, recent time series foundation models (TSFMs) introduce new capabilities such as zero-shot forecasting that have remained underexplored in current systematic benchmarks. In this paper, we address this gap by reviewing and benchmarking 30 state-of-the-art models across three residential energy forecasting tasks, i.e., load, photovoltaic generation, and BESS operation. Our study aims to (i) evaluate the performance of statistical, machine learning, deep learning models, and zero-shot TSFMs, (ii) identify the benefits, limitations, and trade-offs involved in selecting forecasting models for practical applications, and (iii) examine how model performance varies across different forecasting tasks. Using real-world datasets from Belgium, Germany, the Netherlands, and the United Kingdom, we apply a unified data pre-processing pipeline, default model configurations, and multiple temporal sampling rates, and evaluate the performance of these models using mean absolute error and root mean squared error. Our experimental results show that LightGBM provides the most reliable and accurate performance among trained models while Timer-XL demonstrate strong zero-shot forecasting capabilities among TSFMs. The source code is publicly available at https://github.com/dcdube/Forecast-Models-Evaluation .