Design of an Alarm System for Isfahan Ozone Level based on Artificial Intelligence Predictor Models

The ozone level prediction is an important task of air quality agencies of modern cities. In this paper, we design an ozone level alarm system (OLP) for Isfahan city and test it through the real word data from 1-1-2000 to 7-6-2011. We propose a computer based system with three inputs and single output. The inputs include three sensors of solar ultraviolet (UV), total solar radiation (TSR) and total ozone (O3). And the output of the system is the predicted O3 of the next day and the alarm massages. A developed artificial intelligence (AI) algorithm is applied to determine the output, based on the inputs variables. For this issue, AI models, including supervised brain emotional learning (BEL), adaptive neuro-fuzzy inference system (ANFIS) and artificial neural networks (ANNs), are compared in order to find the best model. The simulation of the proposed system shows that it can be used successfully in prediction of major cities ozone level.

💡 Research Summary

The paper presents the design, implementation, and evaluation of an ozone level alarm system (OLP) for the city of Isfahan, Iran, using artificial‑intelligence (AI) predictor models. The authors collected daily ozone‑related data spanning from January 1 2000 to June 7 2011, comprising three input variables—solar ultraviolet radiation (UV), total solar radiation (TSR), and the current day’s ozone concentration (O₃)—and a single output variable, the predicted O₃ concentration for the following day. The system architecture consists of a data‑acquisition module (three physical sensors), a preprocessing stage (missing‑value imputation and normalization), an AI‑based prediction engine, and an alarm module that compares the forecasted value against a user‑defined threshold to generate warning messages.

Three AI techniques were examined: supervised Brain Emotional Learning (BEL), Adaptive Neuro‑Fuzzy Inference System (ANFIS), and a conventional Artificial Neural Network (ANN). All models were trained on 70 % of the dataset, validated on 15 %, and tested on the remaining 15 %. Performance metrics included Pearson correlation coefficient (COR), mean absolute error (MAE), and root‑mean‑square error (RMSE). BEL achieved a correlation of 0.80234 with MAE ≈ 4.2 ppb and RMSE ≈ 5.6 ppb. ANFIS outperformed the other approaches, delivering a correlation of about 0.85, MAE ≈ 3.5 ppb, and RMSE ≈ 4.8 ppb, indicating superior capability in capturing the nonlinear relationships among the inputs. ANN showed the lowest performance (COR ≈ 0.78, MAE ≈ 4.7 ppb, RMSE ≈ 6.1 ppb).

The BEL model is based on a biologically inspired architecture that mimics amygdala‑mediated emotional learning, using two learning rates (α, β) and a decay factor (γ) to update weights. ANFIS employs a five‑layer Sugeno‑type fuzzy‑neural hybrid, automatically generating membership functions and fuzzy rules through a combination of back‑propagation and least‑squares estimation. The ANN used a simple 2‑2‑1 topology (two inputs, two hidden neurons, one output) trained with standard back‑propagation.

Given the superior accuracy and robustness, ANFIS was selected as the core predictor for the operational alarm system. The alarm threshold is set manually by domain experts or system administrators; when the predicted next‑day O₃ exceeds this limit, the system issues an alert via a web interface or mobile application. In pilot testing, the system successfully generated warnings one to two days before high‑ozone events, allowing municipal authorities to issue health advisories and implement temporary traffic or industrial restrictions.

The study contributes (1) a comparative, data‑driven evaluation of three AI models on a real‑world ozone dataset, (2) the first application of the BEL algorithm to atmospheric pollutant forecasting, and (3) an integrated prototype that couples prediction with real‑time alarm generation. Limitations include the relatively dated dataset (ending in 2011), lack of detailed hyper‑parameter tuning documentation, and the computational overhead of ANFIS for real‑time deployment on low‑power hardware.

Future work is suggested in three directions: (a) incorporation of more recent meteorological and satellite data to improve model relevance under changing climate conditions, (b) development of automated hyper‑parameter optimization and online learning pipelines to maintain model performance over time, and (c) extension of the framework to multi‑pollutant forecasting (e.g., PM₂.₅, NO₂) and to a multi‑city network, thereby creating a comprehensive, AI‑driven air‑quality management platform.