The fast development and wide utilization of distributed generations (DGs), such as Photovoltaic panels and wind turbines, provide environmentally friendly renewable energy. However, inappropriate operation, sizing, and placement of DGs could increase the power losses and reduce the stability of the power network. Load forecasting is critical to the electrical utilities to schedule power generation and distribution. In this dissertation, a framework is proposed for load forecasting and optimal operation of power system with DGs in the distribution feeder-level.
In the first part, a nonparametric method, the Bayesian Additive Regression Trees (BART), is introduced for day-ahead peak load forecasting. The detailed correlation analysis of peak load and weather information is performed in a residential area in Albuquerque and a business area in the North Central of Texas for two different two-years periods. Next, the BART method is applied with a principled permutation-based inferential variable selection approach. The BART method’s prediction accuracy is then compared with the Multiple Linear Regression (MLR), the Support Vector Machine (SVM) and the composite kernel of Gaussian Process Regression (GPR). The forecasting results are then measured by Mean Square Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Forecasting Error (MAFE), R2, and Mean Absolute Percentage Error (MAPE). The BART method displays the best prediction accuracy for every index.
In the second part, a new framework of distribution feeder-level short-term and very short-term load forecasting is proposed. First, a composite Mat´ern kernels (CMKs) based Gaussian Process is designed for day-ahead load forecasting based on four years of recorded data and kernels comparison. A data selection algorithm is proposed to improve the prediction further. Second, a three-step daily curve tuning algorithm is designed based on the dictionary learning algorithm, K-SVD, to improve the forecasting results further. In step one, the dictionary is built by using the K-SVD to decompose the output of the CMKs. In step two, for a certain length of atoms, tuned curves are generated by using the K-SVD to learn the known daily load. A curve selection model is designed to choose the best-tuned curve based on the linear regression models with forecasting errors as feedback. In step three, the final tuned curve is selected by the minimization of the mean daily load difference. The framework is verified using two-year private data from the residential area and two-year public data from the business area with three aspects of results.
In the third part, an optimal method to plan and dynamically operate the DG based on the modified nondominated sorting genetic algorithm II (NSGA-II) is proposed. First, the uncertainty of load and DG (photovoltaic panels) output are considered. Second, the placement of a DG is defined by voltage sensitivity analysis. A multi-objective problem is then formulated to find the optimal daily operation of DG. To solve the problem, a fuzzy logic decision model is designed to modify the traditional NSGA II that selects an optimally compromised solution from the Pareto front. Furthermore, to increase the modified NSGA II computation speed, the population initialization space is reduced, and the population is selected and saved for the next generation based on load analysis. With the accurate load forecasting as in part one and two, the initialization space could be reduced further. The method is tested on the IEEE 14 bus, and it is compared with other two optimal methods. The results on reducing the power losses, voltage deviations, and increasing the algorithm speed demonstrate the effectiveness of this method.
peak load forecasting, short-term load forecasting, very short-term load forecasting, multi-objective optimization, distributed generation sizing and placement
Level of Degree
Electrical and Computer Engineering
First Committee Member (Chair)
Dr. Jane M. Lehr
Second Committee Member
Dr. Manel Martínez-Ramón
Third Committee Member
Dr. Olga Lavrova
Fourth Committee Member
Dr. Andrea A. Mammoli
Chen, Tairen. "A Machine Learning Based Framework for Load Forecasting And Optimal Operation of Power Systems with Distributed Generation." (2019). https://digitalrepository.unm.edu/ece_etds/361
Available for download on Monday, July 29, 2019