THE DEVELOPMENT OF REINFORCEMENT LEARNING CONTROLLER FOR AIR CONDITIONING SYSTEM TO INCREASE SELF-CONSUMPTION IN UNIVERSITY BUILDINGS

• Main Author: Mahesa Nanda, Rezky
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/67219
• Institution: Institut Teknologi Bandung

The use of massive solar energy possesses an adverse effect to fossil-fueled plants. To resolve this issue, photovoltaics production must be immediately consumed by building loads. The immediate consumption will be referred to as self-consumption from this point onwards. In addition to that, the building load during the night should be decreased. The fact that air conditioning is the major energy consumer in buildings makes it more interesting to implement load shifting towards its load. However, controlling air conditioning systems in a building is a complicated problem. Especially in university buildings where the room usage pattern and the number of occupants often cannot be accurately predicted. In this research, a reinforcement learning controller was developed to increase university building self-consumption by load shifting without compromising occupant’s thermal comfort. The novelty of this research lies on the definition of the photovoltaic system production constraint in the objective function and the defining of the allowable temperature range based on the typical university buildings occupants’ activity level and clothing insulation, so the control process does not violate the thermal comfort range. The approach that this research used was modeling and simulation. The case study object that was used in this research was Labtek XIV SBM ITB building which is in Bandung, Indonesia. This building was modeled using SketchUp and OpenStudio. In this research, the modeling has been limited to the ground floor and the first floor of the real building for the sake of simplicity. The temperature setpoint of the air conditioning system was set constant at 24°C as an initial value. The model that has been built was then validated. The validation has been done by analyzing the energy consumption pattern of the air conditioning system which was obtained by simulation and measurement. In general, both data exhibit the same pattern, where the peak load tends to happen during afternoon (12:00-18:00 hrs). The daily RMSE error varied from 0.002 to 0.006. Those values suggest that the model’s power intensity will deviate less than 0.006 kW/m2 or equivalent to 17% of the measured maximum power intensity. This deviation is relatively acceptable. The daily Pearson coefficients of correlation indicates strong correlation between the model and the measured data, except for Wednesday. The reinforcement learning algorithm chosen for this research was Proximal Policy Optimization (PPO) because it is stable, simple, and does not need too expensive computation resource. Nevertheless, a hyperparameter tuning, i.e., learning rate, has been done in this research. From the final tuning, the controller with a leaning rate of 0.0001 was chosen because the reward earned by the end of the tuning process is higher than the other. In this research, the controller was evaluated on the 4 weeks when there is equinox and solstices. The parameter evaluated includes the change of the building self-consumption, the change of peak load and bills saving, and the thermal comfort range. It was obtained that the controller was able to increase the building’s average maximum self-consumption of as high as 6.58% and saved 400 kWh annually peak load energy or equal to Rp574,432.00 annually. The thermal comfort range was found to be around ?0.5??????????????0.5 except for the Faculty Lounge room. The mentioned room needs to be conditioned with a higher temperature range to achieve thermal comfort.