Day-Ahead Hourly Forecasting of Power Generation from Photovoltaic Plants (1903.06800v1)

Abstract: The ability to accurately forecast power generation from renewable sources is nowadays recognised as a fundamental skill to improve the operation of power systems. Despite the general interest of the power community in this topic, it is not always simple to compare different forecasting methodologies, and infer the impact of single components in providing accurate predictions. In this paper we extensively compare simple forecasting methodologies with more sophisticated ones over 32 photovoltaic plants of different size and technology over a whole year. Also, we try to evaluate the impact of weather conditions and weather forecasts on the prediction of PV power generation.

• The paper provides an empirical analysis comparing diverse methods for day-ahead hourly forecasting of photovoltaic power generation across 32 different plants in Italy.
• It finds that an ensemble approach combining various techniques consistently outperforms individual forecasting methods for PV power prediction.
• Accurate day-ahead PV forecasts improve grid reliability, and forecast accuracy can be enhanced by integrating more precise meteorological data.

Day-Ahead Hourly Forecasting of Power Generation from Photovoltaic Plants

The paper "Day-Ahead Hourly Forecasting of Power Generation from Photovoltaic Plants" offers an empirical analysis of forecasting methods for photovoltaic (PV) power generation by juxtaposing various methodologies across different plants and conditions. In modern energy systems, orchestrating the stable integration of distributed energy resources (DERs), particularly with their innate variability, is imperative. This paper addresses the challenge by focusing on improving day-ahead predictions for PV plants, which is critical for enhancing grid reliability and operation efficiency.

The research dives into a comprehensive comparison of both rudimentary and advanced forecasting techniques over 32 distinct PV plants in Italy, which encompass a variety of technologies and geographical settings. By doing so, the researchers aim to delineate the impact that various weather conditions and forecast methodologies have on the accuracy of PV power predictions.

Methodologies Assessed

In evaluating forecasting models, the paper employs:

1. Grey-box Model (GB): This simplistic model utilizes a second-order equation correlating irradiance and power generation, relying primarily on historical data for parameter adjustments.
2. Neural Networks (NN): A standard approach employing a static feed-forward structure to interpret the relation between meteorological inputs and power output.
3. k-Nearest Neighbours (kNN): A method utilizing historical similarity to predict power output, implementing a weighted average based on previously analogous meteorological conditions.
4. Quantile Random Forest (QRF): A stochastic method tailored for ensemble predictions, capturing prediction intervals rather than single-point estimates.
5. Support Vector Regression (SVR): An algorithm focusing on minimizing training error within specified bounds, leveraging a Gaussian kernel for input transformation.
6. Ensemble Approach (ENS): A weighted combination of the aforementioned techniques, optimized for overall predictive accuracy through stacked generalization.

Key Findings

The ensemble approach (ENS) consistently outperforms individual methodologies across all measured indices, evidencing the strength of integrating diverse predictive frameworks. Notably, the paper corroborates findings from the Global Energy Forecasting Competition 2014, affirming that ensemble and nonparametric models like QRF and kNN show superior performance, especially for solar forecasting tasks.

Moreover, the impact of weather conditions on forecast accuracy is meticulously evaluated through the Clear Sky Index (CSI), revealing that accurate forecasts are more challenging with intermediate CSI values. This establishes ENS, QRF, and kNN as not only superior but robust under varying meteorological conditions.

Contrary to potential assumptions, some sophisticated methodologies do not overwhelmingly outperform simpler ones in all instances. The Grey-box model, despite its simplicity, provides relatively accurate forecasts, reinforcing the utility of combining basic and advanced methods. The results are further augmented by assessing other performance indices such as nRMSE and nMBE to provide a comprehensive perspective on forecast bias and accuracy dispersion.

Implications and Future Prospects

This research has significant implications for improving the operational efficacy of power systems relying on distributed renewable sources. With accurate day-ahead forecasts, power operators can optimize dispatch planning, enforce compliance, and mitigate the economic penalties associated with mispredicted supply levels.

Looking forward, the accuracy of PV output forecasts can be enhanced by integrating more precise meteorological data, as the paper observes a reduction in forecast error when employing measured irradiance data instead of forecasts. Therefore, improving satellite data resolution and forecasting algorithms are promising avenues for future work.

In conclusion, this paper provides a pivotal comparative paper that elucidates the nuanced performance of different predictive models in solar energy. This work serves as a stepping stone for subsequent research striving for optimal integration of renewable technologies with existing energy networks.