5. 3 Operation, Performance and Maintenance of PV Systems
Summary / Abstract:
The ever-growing amount of data collected by monitoring services provides a ground-breaking opportunity to the field of Machine Learning (ML) and data-driven modelling of the main components in a Photovoltaic (PV) plant is one of the areas that can benefit from the large amounts of data generated in modern PV plants. This paper presents the main steps for the generation of a data-driven model capable of estimating the power output (P) of an inverter in a solar photovoltaic plant under normal conditions. The inverter’s power output will be estimated based on the irradiance on the plane of the array (IPOA) and module’s cell temperature (TMOD); both IPOA and TMOD will be collected from the inverter’s historical data. The procedure has been tested using SCADA data collected by GreenPowerMonitor and its performance has been validated in operation at several selected test sites. On-site collected data are not free of errors; therefore, the procedure for the generation of data-driven models must include data cleaning routines. The objective of this step is to identify those points that do not conform to the inverter’s statistically normal behavior. After said points have been identified, they are removed from the modelling data set. In its present form, the procedure focuses mostly on cleaning data based on P and IPOA; moreover, it also considers out-of-range data, missing data (i.e. nulls or NaNs), and duplicated time stamps. For the present application, a feed forward Neural Network (NN) has been selected to model the inverter’s power output as function of IPOA, TMOD and the sun’s position (elevation () and azimuth (AZ)). Moreover, the model introduces P, IPOA, TMOD, and AZ from previous time steps (or lags) in order to include the inverter’s behavior over time (i.e. persistence). The optimal NN structure (number of neurons and layers) is derived by means of a parametric study. Results show that the optimized NN can accurately estimate the inverter’s power production with low Root Mean Square Error (RMSE) values; RMSE values of 1-2% are consistently achieved among trained models. However, the simulation time required to execute the parametric study makes it impractical if the goal is to generate a large number of models (each of them representing a different inverter). In order to circumvent this drawback, the use of a fixed-size NN has been explored; additional testing demonstrated that the fixed-sized NN’s performance is on par with the optimized model.