Abstract
The development of smart grid has a higher requirement for the real-time data acquisition, processing and association analysis of mass data, big data is one of the core technologies. Accurate prediction of the output of PV (Photovoltaic) plant is of great significance to the safety and stability of power system. The output of PV plant is influenced by the light intensity, temperature and humidity, and there is a complex relationship between factors, there will be a large error when this prediction using LSR(Least Square Regression) analysis. On the basis of electric power data platform, this paper proposes a method for predicting the output of PV plant based on PLSR(Partial Least Squares Regression), this method considers parallel advantage of MapReduce, focuses on the independence of concurrent events, can meet the coverage requirements of power big data attribute dimension and reduction. Through the experimental of data simulation prove that this method is feasible in the treatment of big power data attribute reduction.
Introduction
Predicting the output of solar power plants is a critical challenge for smart grid operators. Due to the intermittent nature of solar energy, which fluctuates based on weather conditions, inaccurate forecasts can threaten the stability of the power grid. The primary difficulty lies in the fact that the factors influencing power output—such as light intensity, temperature, humidity, and wind speed—are highly correlated or "tangled." Traditional statistical methods like Least Square Regression (LSR) struggle with this multicollinearity, often resulting in high error rates that make them unreliable for real-time grid dispatching.
Methodology and Data
To address these complexities, this study implements Partial Least Squares Regression (PLSR). Unlike standard regression, PLSR extracts "principal components" that capture the most predictive information while filtering out noise from overlapping variables.
The model was developed using historical data from the UQ Center PV plant (University of Queensland). The dataset includes five independent variables: the previous day's output ($x_1$), irradiance ($x_2$), temperature ($x_3$), humidity ($x_4$), and wind speed ($x_5$).
The previous day's output ($x_1$) serves as a proxy for the physical state of the plant (e.g., age, tilt, dust). The system is designed to scale within a Hadoop framework using MapReduce, enabling the processing of massive real-time data streams across distributed computing nodes.
Results and Discussion
The performance of the PLSR model was evaluated against a standard LSR baseline using a testing set from January 16 to January 20, 2014.
As shown in the results, the PLSR model significantly outperformed the LSR model. The PLSR relative error remained between -2.24% and -9.76%, whereas the LSR model exhibited much higher errors, ranging from 23.54% to 27.51%. By effectively handling the multicollinearity of environmental factors, the PLSR method proved to be approximately three times more accurate than the traditional approach.
Conclusion
This research demonstrates that PLSR, integrated into a big data architecture, provides a robust solution for PV power prediction. Future work will focus on refining the model by eliminating irrelevant variables and testing the algorithm's performance across different seasons to ensure long-term reliability in diverse weather conditions.
Summary updated