Smart Meter Energy Data from a Controlled Environment

Physical Laboratory Setup for Real Energy Data Generation from a Controlled Environment Setup Description The energy sector consists of a large variety of complex systems and processes. The ecosystem is currently controlled and overseen mainly by expert systems and expert human operators, but the inclusion of AI-assisted solutions is becoming more and more prominent. Such systems can support human operators with various analytics and decisions; therefore, their continuous development remains a high priority in the energy sector. The development of such AI algorithms and solutions is directly connected to the quality of the provided training data. The energy sector generates large amounts of data, most of it connected to smart energy meter data, but unfortunately, the details of this data are not known. Training data for AI development should be as deterministic as possible, meaning that most details should be described and labeled. For this reason, we developed and built a physical laboratory setup that models a transformer station and five houses, connected to the transformer station (Photos in attached img.zip: The front panel includes six smart meters (1+5) and possibilities for one- or three-phase load connection. The back panel includes breakers for each). The transformer station is represented by one smart energy meter, which measures all consumption/generation of the five houses. Each house is represented by a smart energy meter onto which various loads can be connected, ranging from single- to three-phase loads. The setup is totally customizable, as its topology, connection cable lengths, and loads can be freely modified. This allows us to obtain static data of the system. To also obtain time-series data, the smart energy meters acquire measurements of various electrical parameters at a defined interval. These values are being stored in a backend system and can be exported for further analysis. In the current setup, we limited ourselves to capturing instantaneous values for voltage, current, active power, and power factors for each of the three phases. The setup, therefore, enables us to generate real energy data from a controlled environment. After a connection topology, cable lengths, and specific loads are selected and implemented on the actual laboratory setup, data generation can begin. The laboratory setup enables the linking of specific load events, e.g., a specific load being turned ON or OFF, as well as modeling of bad connections. We can define test cases that simulate real-life scenarios and run them on the setup. Each test case runs for a predetermined amount of time and results in a fully described dataset. For example, if we define 5-second reading interval and run the system for 15 minutes, we obtain a total of 12,960 data points among all nodes and phases. Along with the size of the dataset, we also get high variability of the data with different types of loads switching on and off. Thus, the system produces high-quality datasets combining static and dynamic data that can be used in various supervised or unsupervised ML tasks. Dataset Description The topology and data format are described in the dataset folders. Due to a change in data collection, the second dataset is in a different format. The first dataset was generated with a 5-second reading interval, while the second dataset was generated with a 1-second interval. The topology and wire lengths was the same in both dataset generations. IMPORTANT NOTE: We had some issues with the physical laboratory setup which manifested in higher discrepancies between arithmetic sum of individual meters and values of the SUM meter. The figure attached in img.zip displays power discrepancy in percentages through time. Phases L1 and L3 show consistent difference (which comes from combining voltage and current measurement errors), while the discrepancies in phase L2 show greater and less consistent errors. At the time of the dataset upload, we did not yet find resolve the issue.