Development of a Non-parametric Model for Low-cost Pyranometer Based on Artificial Neural Network
Abstract
Low-cost pyranometers are widely used as a viable and robust alternative to effectively measure the amount of solar radiation incident on a given area during a specific time interval. Currently, artificial intelligence tools such as Artificial Neural Networks (ANNs) are frequently employed to integrate characteristics of physical magnitudes incident on these types of renewable systems. In this work, a non-parametric model of a pyranometer based on a Multi-Layer Perceptron (MLP) ANN was developed to simulate the irradiance curve characteristic of a pyranometer during the clear period of the day. This model uses electrical currents from three photodiodes with different spectral sensitivities as input signals to simulate the Global Horizontal Irradiance (GHI) curve measured by a conventional pyranometer. Data were collected from the INOVA SOLAR network at IFPE Campus Pesqueira, Brazil, covering approximately one month with 5-minute sampling intervals during clear-sky periods from 04:43h to 17:58h daily. The dataset was divided into training (85%) and simulation (15%) sets. The ANN inputs include 15 variables such as temporal information (day of the week encoded as one-hot vectors, day, hour, minute, sine and cosine transformations of minute), photodiode readings, and historical pyranometer GHI measurements. The output target is the broadband GHI measured by the pyranometer. The ANN architecture consists of two hidden layers with 1 and 3 neurons respectively, selected through iterative testing to minimize the Mean Squared Error (MSE) using the Levenberg-Marquardt training algorithm. The model achieved a Mean Absolute Percentage Error (MAPE) of 7.98% on the simulation dataset, demonstrating its feasibility. The model is not generative; retraining is recommended if data profiles change significantly. This approach offers a cost-effective alternative to traditional pyranometers with potential applications in solar energy and climatology.
Keywords
Low-cost pyranometers, Artificial neural network, Multilayer Perceptron network, Solar Radiation Measurement, Global Horizontal Irradiance (GHI)