The Living Forecast: Evolving Day-Ahead Predictions into Intraday Reality

TL;DR

This work tackles intraday forecasting in power systems by leveraging fully probabilistic day-ahead forecasts and updating them online as new observations arrive. It builds a CVAE-based day-ahead forecaster that outputs a Gaussian mixture, which can be Bayes-updated in closed form to produce a coherent intraday distribution without retraining. The proposed PDCC covariance and the explicit updating formulas enable consistent point, quantile, and ensemble forecasts while staying computationally efficient for real-time use. Empirical results on Gipuzkoa SM consumption and Utrecht PV generation show consistent improvements across NLL, MAE, CRPS, and RMSE, with near-term gains driven by strong temporal correlations and improved covariance modeling. Overall, the framework provides a theoretically grounded, practical solution for intraday probabilistic forecasting in modern power systems.

Abstract

Accurate intraday forecasts are essential for power system operations, complementing day-ahead forecasts that gradually lose relevance as new information becomes available. This paper introduces a Bayesian updating mechanism that converts fully probabilistic day-ahead forecasts into intraday forecasts without retraining or re-inference. The approach conditions the Gaussian mixture output of a conditional variational autoencoder-based forecaster on observed measurements, yielding an updated distribution for the remaining horizon that preserves its probabilistic structure. This enables consistent point, quantile, and ensemble forecasts while remaining computationally efficient and suitable for real-time applications. Experiments on household electricity consumption and photovoltaic generation datasets demonstrate that the proposed method improves forecast accuracy up to 25% across likelihood-, sample-, quantile-, and point-based metrics. The largest gains occur in time steps with strong temporal correlation to observed data, and the use of pattern dictionary-based covariance structures further enhances performance. The results highlight a theoretically grounded framework for intraday forecasting in modern power systems.

Figures (7)

• Figure 1: Overview of the proposed Bayesian intraday update mechanism. A fully probabilistic day-ahead forecast, represented as a Gaussian mixture model, is obtained from a conditional variational autoencoder (top). Given partial intraday observations, the model applies a closed-form Bayesian update to adjust the remaining-horizon distribution (bottom), enabling consistent quantile, ensemble, and point forecasts without re-training or re-inference.
• Figure 2: The sample-based metrics AE$(T',t)$ and MAE$(T')$ for the Gipzukoa SM dataset obtained by (a) the baseline non-updated model and (b) the intraday model with the proposed update mechanism. The (c) absolute and (d) percentage differences in the metric AE$(T',t)$ between the models reflect the varying improvements that come with the updates.
• Figure 3: Tuning of the number of GMM components for the (a) Gipuzkoa SM and (b) Utrecht PV datasets.
• Figure 4: Performance traces over the Gipuzkoa SM synthetic test set. Comparing models with full covariance matrices (PDCC) against diagonal ones.
• Figure 5: Performance traces over the Utrecht PV synthetic test set. Comparing models with full covariance matrices (PDCC) against diagonal ones.