[Paper Review] Two-Tier Prediction of Solar Power Generation with Limited Sensing Resource

Yübo Wang, Bin Wang+4 more|arXiv (Cornell University)|Aug 11, 2015

Solar Radiation and Photovoltaics15 references3 citations

TL;DR

This paper proposes a two-tier prediction framework for 24-hour solar power generation using only historical power data and limited sensing resources. It combines global-tier forecasting via weighted k-NN or neural networks with local-tier adaptive residual correction, achieving higher accuracy than conventional day-ahead methods on a 35kW UCLA microgrid testbed.

ABSTRACT

This paper considers a typical solar installations scenario with limited sensing resources. In the literature, there exist either day-ahead solar generation prediction methods with limited accuracy, or high accuracy short timescale methods that are not suitable for applications requiring longer term prediction. We propose a two-tier (global-tier and local-tier) prediction method to improve accuracy for long term (24 hour) solar generation prediction using only the historical power data. In global-tier, we examine two popular heuristic methods: weighted k-Nearest Neighbors (k-NN) and Neural Network (NN). In local-tier, the global-tier results are adaptively updated using real-time analytical residual analysis. The proposed method is validated using the UCLA Microgrid with 35kW of solar generation capacity. Experimental results show that the proposed two-tier prediction method achieves higher accuracy compared to day-ahead predictions while providing the same prediction length. The difference in the overall prediction performance using either weighted k-NN based or NN based in the global-tier are carefully discussed and reasoned. Case studies with a typical sunny day and a cloudy day are carried out to demonstrate the effectiveness of the proposed two-tier predictions.

Motivation & Objective

• To improve long-term (24-hour) solar power generation prediction accuracy under limited sensing resources.
• To address the trade-off between accuracy and prediction horizon in existing solar forecasting methods.
• To develop a two-tier framework that leverages historical power data without requiring real-time weather or irradiance sensors.
• To validate the method on a real-world microgrid with 35kW solar capacity.

Proposed method

• The global-tier uses weighted k-Nearest Neighbors (k-NN) or Neural Networks (NN) to generate initial 24-hour forecasts based solely on historical power data.
• The local-tier adaptively corrects the global-tier predictions using real-time analytical residual analysis.
• Residuals between historical forecasts and actual measurements are modeled and updated dynamically to improve short-term accuracy.
• The framework is implemented and tested on a 35kW solar installation at UCLA Microgrid with limited sensing infrastructure.
• Two global-tier models (weighted k-NN and NN) are compared for performance and robustness across different weather conditions.
• The method avoids reliance on external weather data, relying only on historical power measurements and real-time residual feedback.

Experimental results

Research questions

• RQ1 Can a two-tier prediction framework improve 24-hour solar power forecasting accuracy with only historical power data and limited sensing?
• RQ2 How do weighted k-NN and neural network models compare in global-tier performance for long-term solar forecasting?
• RQ3 To what extent does local-tier residual correction enhance prediction accuracy in variable weather conditions?
• RQ4 Is the proposed method effective for both sunny and cloudy days without additional sensor inputs?
• RQ5 Can the framework maintain high accuracy across different solar irradiance patterns using only power measurements?

Key findings

• The two-tier method achieves higher prediction accuracy than standard day-ahead forecasting methods for 24-hour solar generation forecasts.
• The weighted k-NN based global-tier outperforms the NN-based global-tier in terms of stability and accuracy under varying weather conditions.
• The local-tier residual correction significantly improves short-term forecast accuracy by adapting to real-time deviations.
• Case studies on a sunny day and a cloudy day confirm the method's robustness and effectiveness across diverse irradiance patterns.
• The framework maintains high accuracy without requiring additional sensors, relying solely on historical power data and real-time residuals.