Scaling Machine Learning for Localized Climate Risk Modeling

🌍 The Core Challenge: Moving Past the "Persistence Trap"

Standard global climate simulations often fall flat when attempting to predict high-resolution, localized weather transitions in extreme environmental hotspots like Spain’s Mediterranean coast. When training predictive models on this data, algorithms frequently fall into a computational "persistence trap"—essentially guessing tomorrow's conditions by simply duplicating today’s inputs.

This research project explores a rigorous, data-driven methodology to break this predictive plateau. By building robust automated data orchestration pipelines and transitioning from classical statistical baselines to advanced deep learning architectures, this project forces networks to learn the true underlying physical and atmospheric dynamics of climate shifts.

🛠️ The Architecture: Data Engineering & Deep Learning

This independent research was executed across a dual-phased development lifecycle, processing a massive 30-year daily historical meteorological dataset consisting of over 10,950 individual, rate-limited API data extractions.

1. The Classical Baseline Model: Utilizing an Ordinary Least Squares (OLS) Linear Regression node architecture within KNIME, the project first established a structural performance benchmark. While effective at isolating linear correlations, this model exposed the absolute mathematical limits of non-sequential modeling when confronting highly dynamic weather systems, topping out at an R² of 0.582 and a Mean Absolute Error (MAE) of 2.84°C.

2. The Deep Learning Pipeline: To capture complex, non-linear spatio-temporal trends, the more advanced model used a multi-variable Keras Stacked Long Short-Term Memory (LSTM) network. By executing systemic ablation studies, intentionally isolating target variables, and deploying Z-score standardization to strip out deep seasonal data biases, this architecture successfully shattered the persistence trap—elevating final model accuracy to an R² of 0.871 with a sharp MAE of just 1.56°C.

📂 Open-Source Workflows Available on GitHub

To maintain full scientific transparency, reproducibility, and a commitment to open-source climate science, the complete execution frameworks for both phases of this project are fully documented and public. The integrated KNIME workbooks (.knwf), source data, and standalone Python orchestration scripts may be downloaded directly from the link below.

👉 Explore the Complete Project Repositories on GitHub (https://github.com/soapangels)

🚀 What's Next: The Cross-European Thermal Paradox Analysis

As extreme, unprecedented summer heatwaves sweep across Europe in summer 2026, traditional definitions of "climate safety zones" are being actively challenged. Currently, by adapting and deploying this exact API-driven predictive pipeline to compile a comparative analysis, the project is exploring the region of Odense, Denmark, with a goal of exploring the feasibility of relocation to this area.

This upcoming phase will utilize deep learning and anomaly detection to model the Urban Thermal Paradox—analyzing real-feel indoor heat-stress indices against Northern Europe's high-thermal-mass, low-air-conditioning infrastructure.

Follow the website and GitHub profile for upcoming project releases, real-time data visualizations, and publication updates as this comparative European model goes live.