Solar PV forecasting with missing data

• Main Author: Peng, Shuyuan
• Other Authors: Xu Yan
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/152548
• Institution: Nanyang Technological University

This dissertation aims to use missing historical data to build a model that can be used to predict the future of photovoltaic power generation. Since data loss or incomplete data often occurs when using historical data to make PV predictions, it is necessary to train a model and repair the missing data by using highly efficient machine learning algorithms such as "Extreme Learning Machine (ELM)" and "Random Vector Functional Link (RVFL)". These algorithms are randomization-oriented which allows matrix of input weights to be trained in the decision-making, thus improving accuracy in assessment. Firstly, each individual algorithm should be tested. In this part, the parameters of each algorithm need to be adjusted continuously to get the most suitable model. After the parameters were set to have a relatively small error rate, the algorithm was tested in two strategies. The first strategy is to run the algorithm once and record the corresponding prediction error rate. The second strategy is based on the principle of decision-making, which is to train the same learning input for 100 times and record the average value of ensemble results as final error rate. Secondly, the decision-making is made through the hybrid learning method to reduce the error caused by neural network training. According to the parameter settings developed in the last section, two algorithms are combined in the hybrid model. Each algorithm is repeated multiple times, and the total number of repetitions of the two algorithms is 100 times. Then the final hybrid result is analyzed based on the two algorithms through decision-making method. This dissertation mainly optimizes the model through two aspects. The first is to adjust the values of various parameters in the model including activation function, hidden neurons, input features and number of cycles. The second is decision-making. Combining ELM and RVFL with decision-making can decrease the error rate of individual ELM and RVFL. For hybrid forecasting model, decision-making scheme still provides significance to mitigate the negative impact of uncertainty found in the performance of randomized algorithms. Finally, real-time photovoltaic data can be used to predict the subsequent photovoltaic power generation situation.