Environmental sensor measurements and solar-voltage forecasts from a low-cost IoT device: Ulm and Zapopan field deployments

Published: 14 July 2026| Version 2 | DOI: 10.17632/gvkc67gn5x.2
Contributor:
Erick Lara

Description

This dataset contains the raw environmental sensor measurements, on-device model predictions, firmware, and companion mobile application for a low-cost ESP32-based IoT device for environmental monitoring and 24-hour solar-voltage forecasting. It supports the article published in Electronics (2026). Two field deployments are included. A short hardware and firmware validation was carried out in Ulm, Germany. A 115-day campaign was conducted on the rooftop of Building J at the Universidad Autonoma de Guadalajara, Zapopan, Mexico, comprising 84 days of training (6 February to 30 April 2025) and 31 days of autonomous validation (May 2025). All sensor variables (temperature, relative humidity, luminosity as a 0-100 relative index, and solar panel voltage) were recorded at 15-minute intervals. The Zapopan files include the training measurements, the validation-phase sensor inputs, and the corresponding measured-versus-predicted voltage from the embedded feedforward network. The device firmware (.ino) and the Android application (.apk) are also included. See README.txt for a full description of each file.

Files

Steps to reproduce

Sensor data were acquired by the ESP32-based device at 15-minute intervals and logged in JSON on a microSD card, then exported to CSV. A feedforward network (288/10/10/1) was trained offline in Python with TensorFlow/Keras on the Zapopan training file, using min-max normalization with bounds from the training set, the Adam optimizer, and early stopping. The trained weights were deployed to the device as static float32 arrays. During validation the device generated 24-hour-ahead voltage forecasts autonomously, with a daily on-device gradient-descent weight update. The prediction file pairs each measured voltage with the device forecast. Reported error metrics use daytime hours (06:00-19:00) over the 28 clean validation days; the final three days carry an RTC-shifted timestamp and are excluded. Firmware (Devicecode.ino) and the mobile app (Measurements app.apk) are included to reproduce acquisition and monitoring.

Institutions

Categories

Solar Energy, Photovoltaics, Machine Learning, Internet of Things, Sensor

Licence