Power outage prediction with deep learning | |
| Author | Tanawat Benchasirirot |
| Call Number | AIT RSPR no.DSAI-22-01 |
| Subject(s) | Electric power failures Deep learning (Machine learning) |
| Note | A research study submitted in partial fulfillment of the requirements for the degree of Master of Engineering in Data Science and Artificial Intelligence, School of Environment, Resources and Development |
| Publisher | Asian Institute of Technology |
| Series Statement | Research studies project report ; no. DSAI-22-01 |
| Abstract | Power outage occurrence have a large impact to the power industry, users and economy. Mae Sariang is one of the areas in Thailand that face with energy insecurity problem, including power outage occurrences. From the report of Takapong (2018), power outage in Mae Sariang mainly occurred because of the severe weather event. In this research, we aimed to develop a power outage prediction model using deep learning and the model is developed from weather variables. For the dataset, weather datasets are retrieved from two sources; Google Earth Engine (GEE) dataset and Weather Research and Forecasting (WRF) dataset. Both datasets are trained with the machine learning model and are compared to evaluate the quality of the dataset. For power outage data, the load profile data is supported by the Provincial Electricity Authority (PEA). In addition, in terms of the prediction model, we aimed to use Random Forest (RF), Feed-Forward Neural Network (FNN), Bayesian Neural Network (BNN), and Neural Network Ensemble (NNE). After analyzed the data and conducted the experiment by training the model, we found that the important features are temperature and soil water (or moisture), not a gust speed feature like the previous researches. The next finding is that the high-resolution dataset gave a better classification result compared to the low-resolution dataset because the high-resolution can gave a more precise weather value for each area. Furthermore, as our datasets have imbalance classes, we implement the technique to deal with this problem and found that class weighting is more efficient that oversampling technique because the second technique can lead the model overfitting. Lastly, the important finding for model training is that BNN model with class weight tuning gave an impressive classification score, at ~85% of f1-score for validation set and test set. It indicates that this model can classify the power outages efficiently. Interestingly, another result we found is that deep learning models are much efficient compared to the traditional machine learning model. At the same condition, f1-score of deep learning model (BNN) could be reached to 74%, while the traditional machine learning model (RF) has f1-score around 48%. |
| Year | 2022 |
| Corresponding Series Added Entry | Asian Institute of Technology. Research studies project report ; no. DSAI-22-01 |
| Type | Research Study Project Report (RSPR) |
| School | School of Engineering and Technology |
| Department | Department of Information and Communications Technologies (DICT) |
| Academic Program/FoS | Data Science and Artificial Intelligence (DSAI) |
| Chairperson(s) | Mongkol Ekpanyapong; |
| Examination Committee(s) | Dailey, Matthew N.;Singh, Jai Govind; |
| Scholarship Donor(s) | Royal Thai Government Fellowship; |
| Degree | Research Studies Project Report (M. Eng.) - Asian Institute of Technology, 2022 |