DEVELOPMENT OF A PREDICTIVE MODEL FOR MONITORING AND OPTIMIZATION OF CARBON EMISSIONS TO SUPPORT CARBON TRADING DECISION-MAKING (CASE STUDY: PAITON ENERGY COAL-FIRED POWER PLANT)

• Main Author: Ayu Herawati, Neng
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/79651
• Institution: Institut Teknologi Bandung

The Ministry of Energy and Mineral Resources (ESDM) has announced that carbon trading in the first phase will be implemented in 2023, starting with coal-fired power plants (PLTU). Consequently, each power generation unit must calculate greenhouse gas (GHG) emissions. Furthermore, every institution needs to undertake optimization efforts for carbon emissions. PLTU Paiton Energy is one of the coal-fired power plants mandated to follow carbon trading mechanisms as per government regulations. Currently, PLTU Paiton Energy has manually identified GHG emissions but has not yet provided information to support the efficiency of future carbon trading activities. To enhance the efficiency of carbon trading activities, emission predictions can be utilized as an implementation strategy for emission reduction. This research focuses on developing predictive models that provide information on carbon emission predictions, coal fuel consumption, and gross electricity as part of carbon emission optimization efforts. The study uses the CRISP-DM (The Cross Industry Standard Process for Data Mining) methodology to support decision-making in carbon trading activities. Model performance is evaluated using MAE, RMSE, and MAPE metrics. Training involves five machine learning models: Linear Regression, Support Vector Regression, Decision Tree Regression, Random Forest Regression, and LightGBM. The predictive model development employs a machine learning approach combining time series and regression models. Time series models predict each feature or variable as input for regression models to predict carbon emissions. Decision support systems applying rule-oriented decisions guide carbon trading decisions. Testing results for each model at PLTU Paiton Energy Units 3, 7, and 8 identify the Decision Tree Regression Model as the best model for Time Series Method 1 and CEMS Methods. Additionally, the Linear Regression Model performs best in predicting CO2e emissions for Method 1, while the Random Forest Regression Model is optimal for CO2 emissions prediction using the CEMS method. Analysis of testing results at each PLTU Paiton Energy unit yields insights that model quality depends significantly on data characteristics and appropriate model selection. Units with data linearly related to the prediction target show more accurate predictions using the linear regression model. In contrast, units with non- linear data require more complex models like Decision Tree Regression, Random Forest Regression, LightGBM, and Support Vector Regression to address volatility and outliers. Implementing carbon emission prediction models for 2024 results in different carbon trading decisions between Method 1 and CEMS. Method 1 and CEMS decisions for Unit 3 are "selling emissions." However, Method 1 decisions for Units 7 and 8 result in "buying emissions," while CEMS decisions for Units 7 and 8 are "selling emissions." Carbon trading decisions indicate a significant impact from differences in actual emission quantities and data fluctuations in each method during model training, influencing emission prediction results. With the development of predictive models in this research, PLTU Paiton Energy can make informed carbon trading decisions based on emission predictions for future periods. PLTU Paiton Energy can also formulate strategies for carbon trading activities by estimating electricity production outcomes and considering predictions for coal fuel consumption and emissions. This enables PLTU Paiton Energy to identify environmental and financial impacts in the future.