Deterministic and Probabilistic Short-Term Forecasting Approaches for Load Demands and PV Generation

Abstract

As the forecasting algorithm is necessary to provide prosumer’s load profile in active contribution in energy management system this dissertation proposes the deterministic and probabilistic short-term forecasting approaches for load demand and photovoltaic (PV) power generation To provide a computational effective and high accuracy forecasting performance whale optimization algorithm – discrete wavelet transforms – multiple linear regression (WOA-DWT-MLR) are used as short-term load forecasting algorithms tested on both system- and end-user load data sets Like load pattern the PV generation also inherits uncertain behavior and seasonal pattern Thus salp swarm algorithm - convolutional neural network (SSA-CNN) is used to extract the optimal features of PV power generation without time-consuming trial and error in CNN hyperparameter tuning Instead of interpreting the forecasting accuracy only by mean root error (MRE) and mean absolute average percentage (MAPE) an observation about the intended risk that is going to be happened by the forecasting results are studied As the probabilistic approach a prediction interval (PI) with a risk assessment index (RAI) is constructed to comprehend the uncertainty behavior in load demand The PI is modeled by quantile regression random forest (QRRF) complemented by the probability mapping and risk Assessment Index (RAI) Probability mapping is used to capture the hourly distribution of daily load while RAI is used to calculate the risk that might happen tomorrow by having the corresponding forecasting results The proposed methods are tested using data set of Independent System Operator New England (ISO-NE) Shalun office and a 200-kW PV power plant in south Taiwan to represent system load end-user load and PV power generation respectively The simulation results indicate that the proposed short-term deterministic forecasting approaches can effectively improve the forecasting accuracy while the proposed short term probabilistic load forecasting can indirectly quantify the uncertainty that may happen on the next day without extensive computation time

Date of Award	2020
Original language	English
Supervisor	Hong-Tzer Yang (Supervisor)