Machine Learning Implementation at Mölndal Energi AB Evaluating the Possibilities of Implementing Machine Learning within the Production System at Mölndal Energi AB

Sammanfattning

This thesis aimed to evaluate the possibilities of implementing machine learningwithin the production system at Mölndal Energi. The outcome of the thesis includedan overview of potential applications and recommendations for implementation,along with two proof of concepts demonstrating simplified versions of potentialapplications. The aim was to support Mölndal Energi in improving certain areas oftheir production system and give the company a foundation for further developmentand increased competitiveness.The literature review revealed predictive maintenance, demand forecasting, schedulingand energy storage as relevant applications. It was concluded that predictivemaintenance would be the most relevant to implement due to the high potentialof improvement. Mölndal Energi already use, both directly and indirectly throughGöteborg Energi, external machine learning applications for both demand forecastingand scheduling. There are several other areas of interest that could be investigatedfurther in future work, such as the usage of digital twins to simulate anddebug systems, as well as inspection and fault detection using drones and augmentedreality.Two proof of concepts were developed using the programming language Python ina Jupyter Notebook environment. In the first proof of concept, a machine learningmodel was developed to predict the maximum load of a boiler based on operationaldata. It used an LGBM algorithm for quantile regression and the results were evaluatedby comparing the predicted output with actual output for a number of randomsamples, which showed reasonable results. Yet, further testing would be requiredbefore deployment. A second proof of concept was developed to identify deviationswithin fuel data. Three regressor algorithms were compared (LGBM, XGBand RF) and results were evaluated using tables and plots of the deviations. Onceagain results seemed reasonable, however further testing would be required beforeoperational usage. In future work, both proof of concepts could be improved bycomparing more algorithms and fine-tuning the parameters further. The anomalydetection model could also be applied to similar areas by changing the dataset.