Predicting air quality during forest fires represents one of the most complex challenges in environmental modelling. Fine PM2.5 particles, capable of penetrating deep into the lungs, reach extreme concentrations that far exceed the usual safety thresholds. Traditional models, based on recurrent neural networks such as LSTM or BiLSTM, have proven to be effective, but the emergence of foundational time-series models has opened a new debate: can these large pre-trained models, trained with big data from other domains, correctly generalize to extreme phenomena such as wildfire smoke? The results of recent evaluations, which compare models such as TimesFM, Chronos-2, Moirai-2 or Time-MoE against classical architectures, reveal a nuanced reality: recurrent models trained specifically for the task still outperform any foundational model, especially in detecting dangerous peaks.
This finding has profound implications for the development of operating early warning systems. In real situations, it is not enough to predict the average concentration; We must accurately anticipate those moments when air quality becomes extremely harmful. Common metrics such as mean absolute error (MAE) or root mean square error (RMSE) hide poor performance in distribution queues. For example, some foundational models in zero-shot mode showed severe instability in the RMSE for 24-hour horizons, with negative values of R², indicating that they consistently failed in extreme events. Only after applying efficient tuning techniques such as LoRA was it possible to stabilize performance, although without reaching the levels of a BiLSTM trained from scratch with local data.
From a business and technology perspective, this reinforces the importance of having customized solutions and not relying exclusively on generic models. In artificial intelligence for companies, the key is not in the size of the model but in the suitability of the specific problem. Recurrent architectures can be more computationally efficient and perform better against out-of-distribution data, provided that a representative training set is available. For an organization that needs to implement an air quality monitoring system, it is more strategic to invest in custom software development that integrates trained models with local historical data, rather than adopting standard solutions that can fail precisely when they are needed most.
In this context, technology companies such as Q2BSTUDIO offer services that go beyond the simple deployment of models. Creating tailor-made applications for environmental management involves not only selecting the right algorithm, but also orchestrating the data infrastructure, integrating with weather and sensor sources, and visualizing results in real time. AWS and Azure cloud services provide the scalability needed to process large volumes of hourly data from hundreds of monitoring stations, while business intelligence tools such as Power BI allow you to build dashboards that alert civil protection teams intuitively.
Cybersecurity also plays a critical role, especially when early warning systems are connected to government networks or critical infrastructure. An attack that manipulates predictions could have catastrophic consequences. Therefore, any implementation must include robust security protocols, such as those that are integrated into the pentesting and cybersecurity services offered by Q2BSTUDIO. In addition, process automation using AI agents allows the system not only to predict, but also to take corrective actions, such as activating evacuation protocols or modifying the operation of ventilation systems in public buildings.
The aforementioned study shows that foundational models are not a silver bullet. Its main advantage—transferred learning from large data corpora—becomes a weakness when the application domain has distributions that are very different from those of training. Wildfires generate PM2.5 spikes that can be orders of magnitude higher than typical urban pollution, and pre-trained models typically haven't seen those values during their learning phase. Instead, a BiLSTM trained on 12 years of data from 79 stations in California, with more than a thousand fire incidents, learns to recognize precisely those extreme patterns. In fact, the F1 exceedance for the highest threshold (dangerous, >225.5 μg/m³) reached 0.63 compared to 0.54 for the best fine-tuned foundational model.
These results offer practical guidance for any organization that wants to implement environmental prediction systems. The first thing is to have a sufficiently long and representative historical dataset of extreme events. The second thing is not to underestimate the capacity of well-trained classical recurring architectures. Third, consider efficient tuning as an intermediate option when training data is scarce, but always validating performance in distribution queues. And fourthly, to integrate everything into a platform that allows the continuous updating of the models, the monitoring of their performance and automated action.
The application of business intelligence services such as Power BI is especially useful for transforming numerical predictions into operational decisions. A dashboard showing heat maps of predicted concentrations, along with threshold alerts, can be consulted by emergency teams in real time from any device. Artificial intelligence for companies should not be limited to generating numbers; it should facilitate interpretation and action. In this sense, AI agents can act as virtual assistants that suggest evacuation routes, confinement schedules or even coordinate the shipment of respiratory protection supplies to the most affected areas.
In short, the evaluation of foundational models for PM2.5 in forest fires reminds us that technological innovation does not always follow a straight line. Sometimes, more consolidated approaches, combined with careful data engineering and a robust deployment platform, deliver superior results than massive but generic models. For companies looking to develop high-impact solutions, the key is in balance: leveraging the capabilities of artificial intelligence and the cloud, but without losing sight of the specificity of the problem. At Q2BSTUDIO we understand that each client has unique needs, which is why our portfolio ranges from the creation of custom software to the integration of cloud services, ensuring that the solutions are not only technically advanced, but also operationally effective and secure.
