1. Introduction
Understanding weather patterns and their changes is essential to optimize agricultural production and ensure food security [
1,
2], while being closely related to multiple other sectors such as aviation, general transportation [
3], and even influencing local culture [
4]. As such, there has been a historical interest in recording various meteorological variables, such as air temperature, relative humidity, atmospheric pressure, rainfall, wind speed, direction, and many others [
5]. Traditionally, in conventional weather stations (CWS), measurements used to be taken using analog instruments, such as barographs, anemographs, dry and wet-bulb mercury-based thermometers from thermohygrographs, and many others [
5,
6,
7], with the first AWSs (Automatic Weather Stations) emerging only during World War II, developed by the US Navy [
8]. Furthermore, since the 1980s, AWSs have been increasingly used in the agricultural sector due to improvements in the energy efficiency of data loggers and communications [
9]. This phenomenon is also due to the advent of low-power and low-cost wireless modules, which motivated their applications to expand from industrial to agricultural applications [
10], allowing the development of IoT (Internet of Things) in this field. Given such socioeconomic importance, general academic output to monitor climate changes has increased significantly [
11]. Some biomes, such as the Brazilian Cerrado, are known to be getting hotter, drier, and, therefore, more prone to wildfires [
12], which highlights the urge for AWSs in such places.
Considering the agricultural applications of IoT, most technical productions in this field occur in a fragmented manner and lack cooperation [
13], especially for solutions in early stages of development—thus the importance of open architectures and standards in this matter. Besides all advances, most professional AWS can still cost tens of thousands of dollars [
14,
15,
16], imposing challenges to the large-scale deployment and maintenance of such devices [
17], with inexpensive models emerging as a necessity. In this context, real-time meteorological information should be systematically researched, as well as the prediction of renewable energy and estimating the impact of natural disasters [
18]. Furthermore, Stith et al. (2018) [
19] emphasize the importance of meteorological records for predictive practices based on numerical methods. Such data also make the development of ML (Machine Learning) and DL (Deep Learning) methods viable, as highlighted by Da Silva et al. (2024) [
20], Zaytar (2016) [
21], and Kreuzer et al. (2020) [
22]. In these works, LSTM (Long Short-Term Memory) networks are widely applied to generate accurate hourly forecasts, with the first presenting a public solution applied to the Cerrado region of Brasilia.
Thus, this study aims to provide a complete, open-source, and open-circuit low-cost model for microcontrolled solar-powered AWSs that are capable of being integrated into intelligent forecasts and prove themselves accurate for the Cerrado Region of Brasilia. It is intended to monitor several data, such as temperature, humidity, pressure, UV irradiance, ambient illuminance, anemometry, rainfall, air quality, lightning, and hardware health, as well as estimating total solar irradiance. In addition to being WiFi-integrated, its Web API (Application Programming Interface) is desired to run an MZDN-HF (Meteorological Zone Delimited Neural Network–Hourly Forecaster) [
20] compilation to execute intelligent 24 h forecasts and confirm whether or not it is applicable for custom stations in Brasilia without requiring new compilations.
Related Studies
In the context of microcontroller-based AWS, Parvez et al. (2016) [
23] proposes a photovoltaic-powered station based on LM-35, HSM-20G, MPL115A1, GUVA-S12SD, and TEMT6000 modules for temperature, humidity, pressure, UVI (Ultraviolet Index), and illuminance sensing, respectively, along with an LCD (Liquid Crystal Display). In addition to a vast set of air quality sensors, an authorial rainfall gauge and anemometers were proposed. Despite MPL115A1 being discontinued, the prototype is very complete and integrates GSM (Global System for Mobile Communications) communications, focusing on inexpensiveness for popular use in rural areas of Bangladesh. In the next year, Li et al. (2017) [
18] developed a GPRS (General Packet Radio Service)-oriented portable station to measure wind speed, wind direction, light intensity, temperature, and humidity in Jinan, China, with a DS18B20 and a DHT11 for thermohygrometry and a BH1750FVI for luximetry. Although it presented good results, with little data loss during transmissions, it should be noted that some of the sensors have a reduced range, as DHT11 varies from 20% to 80% in humidity. Therefore, although prominent, low-cost, and compact, such a prototype might not be suitable for very humid or dry zones. Posteriorly, Megantoro et al. (2021) proposed a wireless ESP-32-based AWS model [
24], equipped with a DHT11 and BME280 for thermohygrometry, a rainfall gauge, an anemometer, a UVI module, and several MQ-family sensors for air quality in Airlangga (Indonesia). In addition to its WiFi connection, the author’s model had an LCD for immediate data display. Although not solar-powered, it brought a vast diversity of measurements into a compact prototype with modern sensors and public circuitry details.
Furthermore, Wang et al. (2022) [
25] made significant contributions when proposing a solar-powered AWS with open circuit design based on a BME680—a thermo-hygro-barometer plus air quality meter, all in one. That being a more recent work, its sensor choice is both more precise and cheaper than the sets previously reported. BME68X air quality resistance was also paired with and complemented by a ZPH02 dust sensing module. In terms of solar data, the author cites VEML6070 for I
2C UV irradiance readings, despite its obsoletism [
26]. Furthermore, Wang et al. (2022) [
25] highlight the social importance of adapting technology to Chinese reality with a national production of weather station models. Regarding air quality, Bhandekar et al. (2024) [
27] proposed a solar-powered AWS capable of sensing air quality through a set of many MQ sensors, measuring many gases, vapors, alongside with PM (Particulate Matter) specific modules. Similarly, Fahim et al. (2023) [
28] proposed a prototype that monitors air pollutants to infer the AQI (Air Quality Index) through Fuzzy Logic. Although not deployed, nor solar-powered, it presents a significant contribution in the field of A.I. (Artificial Intelligence) integrated AWSs.
Furthermore, in the field of ML-based weather forecasts, an increasing number of research works have been produced, with many relying on LSTMs or Convolutional LSTMs, such as Kreuzer et al. (2020) [
22], Hou et al. (2022) [
29], Qing and Niu (2018) [
30], Ozbek et al. (2022) [
31]. In addition, in the Brazilian Cerrado region, Da Silva et al. (2024) [
20] propose an encoder–decoder format of LSTM networks, trained and optimized using 5 years of hourly data provided by Station A001 [
32], an official INMET (Brazil’s National Institute of Meteorology) AWS located in Brasilia. After some parameter hyper tuning experiments, the model was able to predict a multivariate set of hourly data over a one-day horizon. The model is named MZDN-HL, and after hyperparameter optimization, it achieved a robust performance of 1.32 °C, 7.14%, 0.629 hPa, and 56.7 W/m
2 of MAE (Mean Absolute Error) for temperature, humidity, pressure, and solar irradiance forecasts, besides being a low-computational-cost solution and easy to implement.
Therefore, the relevance of an AWS that combines wireless integration, solar power, data completeness, open architecture, and good cost–precision balance is clear. Additionally, it is notably interesting to rely on modern sensors that are still available to guarantee the experiment’s full reproducibility and adequate for measurement range, keeping it reliable to local conditions (in this case, the Cerrado biome in Brasilia). Finally, for solutions installed in open environments that are wireless and intended to be minimally energy-saving, the LCD modules’ necessity for in-person data checking, as seen in some works, can be discarded, as well as high-power-consuming sensors. Finally, see
Table 1 for comparisons between this paper’s contributions and some prominent works in the field.
4. Discussion
The collected data have shown good correlation with A001, despite their geographical distance and altitude differences, with temperature, humidity, sea-level pressure, and irradiance performing values above 90% for the first three and 85% for the last one. Their MAE, MSE, and average differences were also reasonably low. Finally, rainfall, wind speed, and wind direction had lower correlations and slightly higher errors, which is expected since rain occurs sparsely and the A001 pole is 10 m long (more than twice the length of this prototype). Although a higher pole could be prototyped, PRT would lose in simplicity both in installation and supporting structure. Although rain may be localized in Brasilia, it is worth noting that the time series average of PRT and A001 pluviometry diverged only by 0.04 mm, very little in this variable context.
In general, the quality of the meteorological data presented is satisfactorily high. Data completeness is also a differential, as the referred station counts with UVA/B/C, illuminance, air quality, and lightning monitoring. The UVA–luximetry correlation was that of = 91.9%, also pretty high and accounting for cross-validation purposes. In addition, most lightning was registered close to (when not during) rainy periods, which also corroborates this measurement accuracy. However, it must be noted how close the luximetry peak (110 kLx) was to the sensor maximum range (120 kLx), raising questions about high illuminance measurements reaching a plateau near the sensor maximum on brighter days in the future. In this context, VEML7700 could be protected with a partial light-blocking dome to ensure sensed values would always be lower than maximum and, then, compensated for in software to match a precision lux meter—therefore allowing final results to exceed 120 kLx cap value and preventing information loss on peaks. The air quality resistance values of BME688 were higher during the night, possibly due to the vegetal and anthropogenic liberation of VOCs during daytime. It must be noted that BME688 is also sensitive to smoke; therefore, it is important in the Cerrado biome context, which suffers from natural seasonal and anthropogenic wildfires. Such events are expected to significantly lower the resistance output, although they did not occur in the observed time window.
Additionally,
Section 3.1.1 achieved a good 54% light–sleep ratio, healthy battery voltage, and no data loss (due to flash memory buffering when RSSI was poor), confirming its robust operation.
Regarding MZDN application, the temperature, humidity, and pressure RMSE (1.80 °C, 7.96%, 0.70 hPa) revealed satisfactory performance, close to the results presented by Da Silva et al. (2024) [
20] (1.77 °C, 9.65%, 0.82 hPa). However, irradiance did not perform as well, exhibiting an RMSE of 143.13 W/m
2, about 22% higher than the 117 W/m
2 archived by the mentioned author in the A001 database, possibly due to different PRT-A001 sensor responsiveness (specially in high zenithal angles) combined with inaccuracies inherent in the approximation made by Equation (
3). When analyzing each feature separately and searching for studies in different datasets, results were revealed to also be close to those of Kreuzer et al. (2020) [
22], Hou et al. [
29], Qing and Niu (2018) [
30], and Ozbek et al. (2022) [
31], which include some of the cutting-edge works in the field of hourly weather forecasting on a 1-day horizon. See
Table 5.
Further enhancements are possible in data gathering, both with new sensors whenever further versions are released and with adding new electronic modules for more variables being monitored, like dust or specific gas concentrations, sky images, etc. In addition, the board contains an unused UART bus which, in the schematic, was meant to be integrated with an A7670C 4G module as a WiFi fallback. The designed board also has a socket for an SCD41 sensor, which is a carbon dioxide concentration I2C module. Such a sensor can pair well with the relative air quality resistance of BME688 in future studies. These enhancements are likely to expand the environmental comprehension of urban areas with high pollution levels.
While the system performed well in Brasilia’s Cerrado biome, its accuracy in other regions with different climatic conditions, such as tropical rainforests or arid zones, remains to be validated. Future reproductions of this study could deploy the system in diverse environments to evaluate its generalizability. In this juncture, new stations could be reproduced to expand climate monitoring in Brazilian territory, especially considering the low-cost components on which the design is based. For this purpose, the firmware code and the board schematics were left open for public use. With respect to forecasts, new MZDN-HF compilations can be adapted and applied, as well as expanded and combined with other models to enhance accuracy and cover more variables, such as rainfall, wind speed, direction, etc.
5. Conclusions
The prototype developed demonstrated the feasibility and reliability of a low-cost (about a few hundred USD), open-source, open-circuit, and AI-integrated AWS, meeting all defined objectives. Using modern sensors and photovoltaic power, the system operated autonomously with stable performance and without data loss, even under adverse network conditions.
Meteorological data collected over a three-week period presented strong correlation with the A001-INMET reference station in most metrics, particularly for temperature, humidity, pressure, and irradiance. The MZDN-HF integration enabled accurate 24 h forecasts using data collected in real time by the prototype, without requiring retraining or local recompilation. The MAE and R2 values obtained demonstrate the potential of this integrated approach for practical use in environmental and agro-meteorological contexts.
Furthermore, the proposed system was developed entirely with personal funding from the authors, focusing on its affordability and reproducibility. All hardware schematics, firmware, and collected data have been made openly available to facilitate scientific cooperation and the dissemination of accessible meteorological infrastructure in regions with limited resources.
Thus, because of its low cost and disclosed nature, it represents a step forward in democratizing access to automated weather stations and helping to center technology development toward popular needs. Therefore, it facilitates agricultural and academic production in countries in the third world, such as Brazil, benefiting its peasantry and meteorological researchers through the exchange of technology and mutual collaboration. As a direct consequence, it also represents a step forward in the field of monitoring climatic changes, especially in high-risk biomes such as the Cerrado region.