Forecasting Intraday Power Output by a Set of PV Systems using Recurrent Neural Networks and Physical Covariates

TL;DR

The paper addresses the challenge of intraday PV power forecasting for grid operation by combining a deterministic physical PV performance model with a neural autoregressive framework. It extends DeepAR with a novel positive Gaussian output and uses covariates from the physical model and system descriptors in a single scale-free model to forecast hourly PV output for a geographically distributed fleet. An extensive ablation study shows the best variant—physical covariates plus system-ID covariate and a positive Gaussian output—achieves a $15.72\%$ skill improvement over the physical baseline, with competitive nRMSE, nMAE, and CRPS metrics, while plain neural models underperform the hybrid approach. The approach provides improved, probabilistic intraday forecasts that can support DSOs in grid operation and risk management, and it offers a scalable framework for incorporating physical model outputs and local system characteristics into time-series forecasting.

Abstract

Accurate intraday forecasts of the power output by PhotoVoltaic (PV) systems are critical to improve the operation of energy distribution grids. We describe a neural autoregressive model that aims to perform such intraday forecasts. We build upon a physical, deterministic PV performance model, the output of which is used as covariates in the context of the neural model. In addition, our application data relates to a geographically distributed set of PV systems. We address all PV sites with a single neural model, which embeds the information about the PV site in specific covariates. We use a scale-free approach which relies on the explicit modeling of seasonal effects. Our proposal repurposes a model initially used in the retail sector and discloses a novel truncated Gaussian output distribution. An ablation study and a comparison to alternative architectures from the literature shows that the components in the best performing proposed model variant work synergistically to reach a skill score of 15.72% with respect to the physical model, used as a baseline.

Figures (7)

• Figure 1: Distinction between regular and rolling forecasts. $t_0$ denotes the present time step in the context of a given data sample.
• Figure 2: Illustration of the DeepAR model. Observed variables are represented as shaded boxes, and latent variables as blank boxes. For the context interval, $z$ variables are always known. For the forecasting interval, the model behaves differently at training and test time. At test time, $\tilde{z}$ variables are sampled according to $p$, forming sample paths. Plain lines represent dependencies between random variables, and the dashed line highlights the reinjected sample.
• Figure 3: l.h.s.: Proportion of missing values per system. r.h.s.: Proportion of missing data per associated month.
• Figure 4: a) nMAE results for variable numbers of positive Gaussian components with a hidden layer size of 100. b) nMAE results for variable hidden layer sizes with 5 components. c) Refinement of the number of components with a hidden layer size of 40.
• Figure 5: Illustration of the relationship between nMAE and nRMSE metrics (a), also magnified after excluding models with nMAE greater than 9% from the representation (b). $\phi$ stands for the baseline physical model. Red lines are the result of a linear regression of the points in the graph, reported along with Pearson and Spearman correlation coefficients.