Deep learning assisted energy management in microgrids

One of the most important requirements of traditional electric energy systems is to keep demand and supply balanced at all times. Although large power plants are used to achieve this target, the integration of renewable energy plants into the utility grid and the uncertainty of the energy generation of these plants jeopardizes this balance. Microgrids can achieve this balance locally and reduce the stress on the distribution network by operating the load and distributed generation plants in a specific area in a coordinated manner. In this thesis, a two-level Energy Management Strategy (EMS) is proposed for optimum operation of a residential microgrid. The most important component of the developed EMS is the Non-Intrusive Load Monitoring (NILM) module that can monitor consumers in the microgrid and instantly detect the appliances they use. NILM is the process of disaggregating the electricity consumption data of end-users measured by a main meter into its appliance-level components using various signal processing methods. The instant identification of actively operating appliances and the estimation of their energy consumption were made with two different deep learning models. A real-time load identification analysis was performed using a Gated Recurrent Units (GRU) based deep learning model for nine different appliances located in the IoT Microgrid Laboratory, Department of Energy Technology, Aalborg University. Depending on the type of appliance analyzed, an analysis success between 65% and 96% was achieved. Apart from this, the difference in accuracy between offline and real-time analysis was tested and it was observed that the accuracy of real-time analysis was 5% to 10% lower. Considering that only load identification analysis is insufficient for the proposed EMS, a new GRU-based deep learning model has been proposed, and both the on/off state of the appliances and their power consumption are analyzed offline. The proposed model performed better than the deep learning models used previously in the literature, as it has the capacity to perform two separate analyzes simultaneously. In the second part of the thesis, an effective EMS mechanism is designed by using the output of NILM analysis. The main purpose of including NILM into the EMS is to reduce the electricity bills of customers by taking into account the consumption habits of them, to benefit from the demand-side management applications offered by the utility grid and to get extra incentives. Each consumer has a different lifestyle and therefore, they have different consumption habits. Thanks to the NILM analysis, the life habits and consumption behaviors of each customer can be learned and a consumer-specific energy management can be designed. In the first level of the two-level EMS, the results obtained from the NILM analysis were statistically analyzed and a consumer-specific cost optimization was performed using the data obtained. The optimum cost is provided by optimally shifting the usage time of the appliances from the periods when electricity prices are high to the periods when the electricity prices are low. In this way, electricity bills were automatically reduced by considering the comfort of consumers. In the second level, it is aimed to operate the microgrid in an optimum manner, taking into account the capacities and constraints of the generation and consumption units existing in the microgrid. In this context, by defining a multi-objective function, it is aimed to reduce the operation cost of the microgrid and to prevent new peaks to be occurred in the utility grid. The developed EMS algorithm has been tested in real-time at Aalborg University, Department of Energy Technology, AC/DC Microgrid Laboratory. The obtained results proved the applicability of the developed EMS algorithm. In addition, the effect of different optimization periods on EMS performance was analyzed. It has been observed that if the estimations used for optimization are of acceptable accuracy, the extension of the optimization period increases the performance of the EMS. As a result of the experiments conducted for 6, 12 and 24 hour periods, it has been observed that the 24 hour optimization period enables the battery to be used more effectively and the microgrid to operate more economically.