Mid-term Electricity Load Forecasting In Cambodia

For electrical system planning,safe operation,and maintenance,mid-term load forecasting(MTLF)is very crucial in a developing country such as Cambodia.Many types of research for MLTF have been established over the last decade.MLTF’s research on Cambodia’s electricity grid,on the other hand,is insufficient to fulfill the optimal planning for rapid demand growth.Meanwhile,the speedy growth of the economy has caused the Cambodia power sectors could not smoothly adapt their process to satisfy the quick growth of demand.Moreover,due to the weather characteristic,during the dry season mostly every year,Cambodia has met the lack of supply situation and system overload while recently there was a whole country power blackout.But when arriving in the rainy season,the supply turns to be more than the demand which causes the waste on energy production.These problems have been considered as the unbalancing of supply and demand.Besides,Cambodia is a developing country which means everything vision is under development the same the power sector which there is much work in progressing and planned to be implemented.However,we noted the specific research on mid-term load forecasting is insufficient in Cambodia country especially the research on load predicting by considering the weather factor impact.As a result,Cambodia will need additional research contributions in this area.This master’s thesis presents the research progress on the mid-term load forecast for Cambodia country.This thesis addressed the general and fundamental aspects of the forecasting process,the utility’s necessity prediction models,and the factors that influence performance analysis,the load forecast techniques and methodologies,and finally discussed several case studies for mid-term horizon by proposing the machine learning models and methodologies to apply on Cambodia’s transmission system load.The developed models for the case studies were Multiple Linear Regression(MLR),Neural Network(NN),Seasonal Autoregressive Moving Average(SARIMA),and Gaussian Process Regression(GPR),and combination hybrid model of SARIMA and GPR which all of these models were practice on monthly index historical datasets of Cambodia’s transmission system load for 10-years historical data from 2010 to 2019.The case studies were implemented into three cases which the first case was using three multivariate models,Multiple Linear Regression(MLR),Gaussian Process Regression(GPR),and Neural Network(NN).This case studies aimed to investigate the effects of time,seasonality,weather,and holidays factor on modeling Cambodia’s system load which the detail of modeling and methodologies were presented in Chapter III and IV.The second case study was the univariate model,in which the SARIMA model was preferred.This case study was not included the affected variable consideration,we only used the system load itself to implement the modeling and forecasting.After finishing the two case studies mentioned above,we attempted to proposed various combinations from the above-discussed model,after calibrating some hybrid model combinations,the hybrid SARIMA and GPR methods have been developed to be discussed in this project.The proposed hybrid technique protocol consisted of two main steps,first SARIMA was trained from the system load variables in training datasets(2010 to 2018 datasets)then we simulated the residual of the model performance(2010-2018)and predicted the future system load in 2019.After that,we proposed a new independent variable by summing the received residuals with system loadvalue in the training datasets while the predicted load in 2019 from SARIMA was as well was added into the testing datasets(2019 datasets)as new variables then we received new training and testing datasets.In the second step,we attempted to train kernel-based GPR on these new training datasets and then simulated the results and evaluated the performance.The details of modeling methodologies were described in chapter IV.The mentioned datasets consisted of five independent variables such as dates,months of the years,relative humidity,wet-bulb temperature,and total nonworking days of each month,and one dependent variable which was Cambodia’s transmission system load.Furthermore,the Mean Absolute Error(MAE)and Mean Absolute Percentage Error(MAPE)are the indicators used to determine the accuracy of predicted values as compared to actual data.The simulated results have revealed that the developed methodology SARIMA-GPR and the NN model outperformed the other models while MAPE errors NN was 1.83 percent and 1.95 percent of SARIMA-GPR,as well as MAE error of NN,was31.13 MW and SARIMA-GPR was 33.3.MW.Besides,the LR model performance is the poorest one among the developed methods when MAE and MAPE of its error were 239.70 MW and 13.93%.while the results from SARIMA and GPR both are acceptable for further discussion.Both models performed with the MAPE of 3.46% the same and a little bit different in of MAE error where 57.77 MW error of SARIMA model and GPR model was59.73 MW.Hence,as a feasible practical utility of the project,we could apply the predicted load as a tool for optimizing planning purposes as well as for the system operator to determine the expectation for the following years.