GENERATIVE ENERGY FORECAST: A HYBRID FORECASTING MODEL COMBINING STATISTICAL AND GENERATIVE AI FOR SHORTTERM ENERGY DEMAND

In the face of rising global energy demands, climate imperatives, and the increasing digitalisation of infrastructure, short-term energy forecasting has become essential for operational efficiency and informed decision-making. While recent advances in Artificial Intelligence (AI), particularly Generative AI (Gen-AI), have expanded the forecasting toolkit, they also introduce new challenges related to interpretability, computational cost, and sustainability. This research aims to develop and validate a novel hybrid forecasting model, Generative Energy Forecast (GenEneCast), that integrates classical time-series decomposition, a Deep Learning Model (DLM), and Gen-AI to improve short-term energy consumption forecasts’ accuracy, interpretability, and sustainability. The study employed a multi-method research design, beginning with a Scoping Review to map the main Gen-AI technologies in energy management. A questionnaire with domain experts was conducted to assess the relevance of these Gen-AI models for energy efficiency, followed by Friedman and Nemenyi statistical tests to rank the most suitable approaches. Based on these findings, the GenEneCast model was developed, integrating Holt-Winters decomposition, LSTM, and GPT-4-Turbo. Beyond narrative generation, the Large Language Model was employed to suggest optimal configurations for both the Holt-Winters model and the LSTM architecture, dynamically adjusting to the statistical profile of each input series. The model was evaluated using realworld energy data through an Action Research approach. The GenEneCast model was assessed through the Model Confidence Set alongside classical and neural forecasting methods to identify statistically superior models within a multiple comparisons framework.The GenEneCast model outperformed traditional standalone methods, enhancing interpretability and model adaptability. The Green Paradox-AI emerged as a key theoretical contribution, highlighting that excessive gains in predictive precision may lead to disproportionate environmental costs. This study underscores the value of integrating decomposition techniques, residual learning, and generative modelling within a unified and adaptive forecasting framework guided by expert input and robust statistical validation. GenEneCast provides practical flexibility, theoretical advancement in hybrid Gen-AI, and strategic value for decisionmakers seeking explainable and resource-aware forecasting solutions.

Keywords: Hybrid Forecasting Model; Generative Artificial Intelligence; Energy Consumption Prediction; Green Paradox-AI; Neural Networks.