Conference paper · 2025
Hyperlocal Temperature and Humidity Prediction Using Supervised Machine Learning
2025 2nd Global AI Summit – International Conference on Artificial Intelligence and Emerging Technology (AI Summit), pp. 940–945, IEEE · Published
Summary
Gigov et al. (2025) compared linear, polynomial, decision tree and random forest regression for hyperlocal prediction of daily average temperature and relative humidity, using ten years of daily Open-Meteo data for Sofia, Berlin and Tokyo. Linear regression and random forest were statistically equivalent for temperature, but linear regression ran faster. For humidity, linear regression was clearly best (R² 0.940–0.964) ahead of random forest (R² 0.913–0.933), while decision trees performed worst (R² 0.741–0.814).
Practical implication. Precision agriculture, smart city infrastructure, event planning and resource-constrained IoT deployments that need real-time predictions can use fast, accurate linear models for hyperlocal temperature and humidity forecasting.
- The Problem
- City-level (hyperlocal) weather prediction needs models that balance accuracy against computational cost, but simple and complex supervised regressors are rarely compared on long daily records.
- The Methodology
- Comparative supervised regression experiment (linear, polynomial, decision tree, random forest) on ten years of daily Open-Meteo data for Sofia, Berlin and Tokyo with a chronological 7:3 split, grid-search tuning, and R², MAE, RMSE and execution-time evaluation.
- The Core Finding
- On ten years of daily data for Sofia, Berlin and Tokyo, linear regression matched random forest for temperature (R² > 0.994) at a fraction of the run time and clearly outperformed it for humidity (R² 0.940–0.964), while decision trees performed worst. Simple linear models are the efficient choice for hyperlocal forecasting.
- The Citation
- Gigov, L., Vajjhala, N. R., & Stoilov, A. (2025). Hyperlocal Temperature and Humidity Prediction Using Supervised Machine Learning. In 2025 2nd Global AI Summit – International Conference on Artificial Intelligence and Emerging Technology (AI Summit) (pp. 940–945). IEEE. https://doi.org/10.1109/AISummit66170.2025.11411099
The summary, key findings, methods, implications and relevance notes on this page are this website’s own description of the paper, written from the published abstract and text. The official abstract and citation details are given further down.
Key findings
- Finding 1. The study trained linear, polynomial, decision tree and random forest regression models on ten years of daily weather data (1 July 2015 to 1 July 2025) for Sofia, Berlin and Tokyo, using a chronological 7:3 train/test split.
- Finding 2. For Sofia temperature prediction, linear regression achieved R² = 0.994, MAE = 0.510 and RMSE = 0.648 on the test set.
- Finding 3. For Berlin temperature prediction, linear regression (R² = 0.996, RMSE = 0.508) slightly outperformed random forest (R² = 0.995, RMSE = 0.527), while random forest took about 0.021 s versus 0.001 s.
- Finding 4. For Tokyo temperature prediction, linear regression and random forest both reached R² = 0.997, with linear regression achieving the lowest MAE (0.325) and RMSE (0.445).
- Finding 5. For relative humidity, linear regression achieved R² of 0.940 (Sofia), 0.959 (Berlin) and 0.964 (Tokyo), outperforming random forest (0.913, 0.933 and 0.923) and decision trees (0.785, 0.814 and 0.741).
- Finding 6. Polynomial regression reached R² ≥ 0.9 for temperature in all three cities, well above the R² ≈ 0.5 reported in earlier temperature forecasting studies, though Tokyo humidity required a 4th-order polynomial.
- Finding 7. The authors conclude that ensemble methods do not guarantee improved accuracy and that model selection should prioritize the relationship structure in the data rather than algorithmic complexity.
Source: Gigov et al. (2025), 2025 2nd Global AI Summit – International Conference on Artificial Intelligence and Emerging Technology (AI Summit), pp. 940–945, IEEE. DOI: 10.1109/AISummit66170.2025.11411099
The study in detail
What question does this paper answer?
Which supervised machine learning regression model – linear, polynomial, decision tree or random forest – best predicts daily average temperature and relative humidity at a hyperlocal (city) level, balancing accuracy and computational cost?
What did the study find, in detail?
Using ten years of daily Open-Meteo data (July 2015–July 2025) for Sofia, Berlin and Tokyo with a chronological 7:3 train/test split, linear regression and random forest were statistically equivalent for temperature (R² > 0.994), but linear regression ran faster (0.001 s vs 0.022–0.040 s). For humidity, linear regression was clearly best (R² 0.940–0.964) ahead of random forest (R² 0.913–0.933), while decision trees performed worst (R² 0.741–0.814).
Why does it matter?
The results challenge the assumption that complex ensemble methods are necessary for accurate weather prediction: model selection should prioritize the relationship structure in the data rather than algorithmic complexity. Fast, accurate linear models suit precision agriculture, smart city infrastructure, event planning and resource-constrained IoT deployments where real-time predictions are essential.
Methods and data
| Research question | Which regression model best predicts hyperlocal average temperature and relative humidity? |
|---|---|
| Design | Comparative experiment of four supervised regression models across three climatically diverse cities |
| Data | Daily historical weather data, 1 July 2015 to 1 July 2025, from Open-Meteo's Historical Weather API for Sofia, Berlin and Tokyo |
| Methods | Linear, polynomial, decision tree and random forest regression; forward fill, Z-score outlier removal, min-max scaling; 7:3 chronological split; grid-search tuning; R², MAE, RMSE and execution time |
| Main result | Linear regression matched random forest for temperature (R² 0.994–0.997) and beat it for humidity (R² 0.940–0.964 vs 0.913–0.933) while running faster |
| Implication | Simpler models can outperform ensembles when relationships are largely linear, favoring efficient real-time deployment |
| Citation | Gigov et al. (2025) · DOI 10.1109/AISummit66170.2025.11411099 |
Standards, datasets and methods named in this study
- Open-Meteo Historical Weather API Data sourceSource of the ten years of daily weather data for Sofia, Berlin and Tokyo.Open-Meteo. Historical Weather API.
Definitions and key terms
- Hyperlocal weather prediction
- Forecasting weather conditions at a fine spatial scale such as a city or community, rather than the coarse grids (e.g. 15 km × 15 km) used by global forecasting systems. — as used in Gigov et al. (2025)
- Random forest regression
- An ensemble of decision trees trained on bootstrapped samples and random feature subsets whose predictions are averaged to give the final prediction. — as used in Gigov et al. (2025)
- Coefficient of determination (R²)
- A measure of how much variance in the target is explained by the model, where values close to 1 indicate a near-perfect fit. — as used in Gigov et al. (2025)
Abstract
Machine learning offers computationally efficient alternatives to traditional numerical weather prediction (NWP) models and global forecasting systems (GFS), which struggle with high computational costs and coarse spatial resolution. This study focuses on hyperlocal weather prediction, comparing linear regression (LR), polynomial regression (PR), decision tree regression (DTR), and random forest regression (RFR) for forecasting average temperature and relative humidity. Using ten years of historical data (2015-2025) from Sofia, Berlin, and Tokyo, we achieved substantial predictive accuracy across climatically diverse regions. For temperature prediction, linear regression and random forest demonstrated statistically equivalent performance (R² > 0.994), with linear regression achieving faster execution times (0.001s vs 0.022-0.040s). For humidity prediction, linear regression demonstrated clear superiority (R² > 0.940) over random forest (R² 0.913-0.933), while decision trees showed the poorest performance (R² 0.741-0.814). Decision trees performed competitively for temperature (R² 0.992-0.993) but struggled significantly with humidity's non-linear dynamics. These results demonstrate that model selection should prioritize the relationship structure in data rather than algorithmic complexity, with linear regression offering a balance of accuracy and computational efficiency. The findings of this study challenge the assumption that complex ensemble methods are necessary for accurate weather prediction, offering practical applications in precision agriculture, smart city infrastructure, event planning, and resource-constrained IoT deployments where real-time predictions are essential.
Abstract as published in 2025 2nd Global AI Summit – International Conference on Artificial Intelligence and Emerging Technology (AI Summit).
Keywords: artificial intelligence; machine learning; weather prediction; regression analysis; hyperlocal weather forecasting; temperature prediction; humidity prediction
Limitations
- Models were validated only with a standard 7:3 chronological train/test split; cross-validation (e.g. training on 2015–2018 and testing on 2024–2025) would improve robustness.
- No exploratory data analysis of feature and target distributions was performed to ensure optimal feature selection.
- Only four algorithms were tested; gradient boosted trees, k-nearest neighbors, LSTM and RNNs remain to be explored, along with real-time API-based input and in-depth error analysis by season or extreme events.
Practical implications
- Resource-constrained IoT deployments and other real-time applications can use linear regression for temperature, since on data for Sofia, Berlin and Tokyo it was statistically equivalent to random forest and ran faster.
- Those selecting models for hyperlocal weather prediction should prioritize the relationship structure in the data rather than algorithmic complexity, since ensemble methods do not guarantee improved accuracy.
- Those forecasting relative humidity can choose linear regression over random forest: across Sofia, Berlin and Tokyo it was clearly best (R² 0.940–0.964), ahead of random forest (R² 0.913–0.933).
- Those applying the study’s results should note that the models were validated only with a standard 7:3 chronological train/test split, without cross-validation.
Drawn from Gigov et al. (2025), 2025 2nd Global AI Summit – International Conference on Artificial Intelligence and Emerging Technology (AI Summit), pp. 940–945, IEEE.
When this research may be relevant
This article may be relevant to literature on hyperlocal or city-level weather prediction with machine learning, temperature and humidity regression, model selection trade-offs between accuracy and computational cost, and lightweight forecasting for IoT, smart agriculture and real-time applications.
Research topics addressed: machine learning; supervised learning; weather forecasting; hyperlocal weather prediction; temperature prediction; relative humidity prediction; linear regression; polynomial regression; decision tree regression; random forest regression; Open-Meteo historical data; Sofia, Berlin and Tokyo; precision agriculture; IoT
How to cite
- APA 7
- Gigov, L., Vajjhala, N. R., & Stoilov, A. (2025). Hyperlocal Temperature and Humidity Prediction Using Supervised Machine Learning. In 2025 2nd Global AI Summit – International Conference on Artificial Intelligence and Emerging Technology (AI Summit) (pp. 940–945). IEEE. https://doi.org/10.1109/AISummit66170.2025.11411099
- MLA 9
- Gigov, Lyuboslav, et al. “Hyperlocal Temperature and Humidity Prediction Using Supervised Machine Learning.” 2025 2nd Global AI Summit – International Conference on Artificial Intelligence and Emerging Technology (AI Summit), IEEE, 2025, pp. 940–45. https://doi.org/10.1109/AISummit66170.2025.11411099.
- IEEE
- L. Gigov, N. R. Vajjhala, and A. Stoilov, “Hyperlocal Temperature and Humidity Prediction Using Supervised Machine Learning,” in 2025 2nd Global AI Summit – International Conference on Artificial Intelligence and Emerging Technology (AI Summit), 2025, pp. 940–945, doi: 10.1109/AISummit66170.2025.11411099.
BibTeX
@inproceedings{gigov2025hyperlocal,
title = {Hyperlocal Temperature and Humidity Prediction Using Supervised Machine Learning},
author = {Gigov, Lyuboslav and Vajjhala, Narasimha Rao and Stoilov, Anton},
booktitle = {2025 2nd Global AI Summit – International Conference on Artificial Intelligence and Emerging Technology (AI Summit)},
pages = {940--945},
year = {2025},
publisher = {IEEE},
doi = {10.1109/AISummit66170.2025.11411099},
url = {https://doi.org/10.1109/AISummit66170.2025.11411099}
}RIS
TY - CPAPER TI - Hyperlocal Temperature and Humidity Prediction Using Supervised Machine Learning AU - Gigov, Lyuboslav AU - Vajjhala, Narasimha Rao AU - Stoilov, Anton T2 - 2025 2nd Global AI Summit – International Conference on Artificial Intelligence and Emerging Technology (AI Summit) PY - 2025 DA - 2025/11/19 SP - 940 EP - 945 PB - IEEE SN - 979-8-3315-5379-1 DO - 10.1109/AISummit66170.2025.11411099 UR - https://www.narasimharao.net/research/hyperlocal-weather-temperature-humidity-prediction-machine-learning/ LA - en ER -
Links and files
- DOI: https://doi.org/10.1109/AISummit66170.2025.11411099
- This page as text: https://www.narasimharao.net/research/hyperlocal-weather-temperature-humidity-prediction-machine-learning/paper.md
- Permanent address of this page: https://www.narasimharao.net/research/hyperlocal-weather-temperature-humidity-prediction-machine-learning/
Related research
Related publications
Fine-tuned DINOv2 and EfficientNet-B0 performed best on nine-class DeepWeeds data, but differences among several high-performing models were not consistently statistically significant across random seeds.
Among 41 Albanian SMEs, data-analytics use correlated positively with operational efficiency, process optimisation, profitability and market growth (r = 0.540 to 0.769), with a significant multivariate effect (MANOVA F(1, 38) = 20.029, p < 0.001); a random forest model explained the outcomes best (R² = 0.661).
Across 92 primary studies (2016–2025), Diffusion Transformers and Gaussian Splatting account for about 36% of studies dated 2024 or later while GAN use has fallen below 3%; training data are demographically skewed and under-documented, and the dominant metrics (PSNR, SSIM, SyncNet) do not measure what deployment requires.
Across 439 successful US government IT projects with 1,230 features, linear regression outperformed random forest and SVM in predicting project manager commitment tenure.
A prefetching approach that builds an online navigation graph from server logs and clusters navigation behaviour with WEKA outperformed conventional prefetching techniques on domain client-group logs, particularly once server idle time allowed prefetching across all user predictions.