The issue of smog in smart cities (SCs) poses serious health risks due to the rising concentrations of air pollutants (APs), including particulate matter, carbon, and sulfur. Their complex and dynamic behavior makes data-driven analysis challenging. Thus, accurate source identification and forecasting of those APs are vital for assessing psychotic disorder risks in hospitalized patients. To address this, a model namedPsych-Airhas been proposed. The proposedPsych-Airmodel employs federated learning (FL) with a customized multivariate bidirectional GRU (BGRU) architecture featuring cross-variable gated attention and parallel temporal encoders. Unlike conventional FL schemes,Psych-Airincorporates heterogeneity-aware and pollutant-sensitive federated optimization, enabling stable and communication-efficient learning across non-IID smart city environments. It also models pollutant and meteorological time series as an interdependent tensor stream, with dynamic gating and synchronized bidirectional states capturing pollutant-specific patterns. A shared fusion layer learns spatiotemporal signatures predictive of psychotic disorder onset while operating securely in a decentralized environment.Psych-Aireffectively isolates latent triggers from noisy environmental data, achieving clinically relevant forecasting. FL-based BGRU identifies key pollutant sources, accounting for 46% of Air Quality Index impact in SCs. The model outperforms traditional ML, DL, and FL methods by 20%, 15%, and 10%, respectively, and supports the assessment of psychotic disorders through statistical analysis. Additionally, Psych-Air promotes sustainable, energy-efficient smart cities by reducing annual costs and CO emissions, proving adaptable across various urban settings.

During the winter season in India, the AQI experiences a decrease due to the limited dispersion of APs caused by MFs. Therefore, we developed a sophisticated green predictive model GAP, which utilizes our designed green technique and a customized big dataset. This dataset is derived from weather research and tailored to forecast future AQI levels in the Indian subcontinent during winter. This dataset has been meticulously curated by amalgamating samples of APs and MFs concentrations, further adjusted to reflect the yearly activity data across various Indian states. The dataset reveals an amplified national emissions rate for$\boldsymbol {PM_{2.5}}$,$\boldsymbol {NO_{2}}$, and$\boldsymbol {CO}$pollutants, exhibiting an increase of 3.6%, 1.3%, and 2.5% in gigagrams per day. ML/DL regressors are then applied to this dataset, with the most effective ML/DL regressors being selected based on their performance. Our paper encompasses an exhaustive examination of existing literature within the realm of air pollution epidemiology. The evaluation results demonstrate that the prediction accuracy of GAP when utilizing LSTM, CNN, MLP, and RNN achieve accuracies of 98.53%, 95.9222%, 96.1555%, and 97.344% in predicting the$\boldsymbol {PM_{2.5}}$,$\boldsymbol {NO_{2}}$, and$\boldsymbol {CO}$concentrations. In contrast, RF, KNN, and SVR yield lower accuracies of 92.511%, 90.333%, and 93.566% for the same AQIs.

Nowadays, Internet of Things (IoT) has become very popular due to its applications in various fields such as industry, commerce, and education. Cities become smart cities by utilizing lots of applications and services of IoT. However, these intelligent applications and services significantly threaten the environment regarding air pollution. Therefore, high accuracy in air pollution monitoring and future air quality predictions have become our primary concern to save human beings from health issues coming from air pollution. In general, deep learning (DL) and federated learning (FL) techniques are suitable for solving various forecasting problems and dealing with the high volatile air components in heterogeneous big data scenarios. This ambiance of DL and FL motivates us to exploit the DL-based Bidirectional Gated Recurrent Unit (BGRU) method for future air quality prediction using big data and federated learning (FL) to train our model in a distributed, decentral, and secure ways. This paper proposes a novel distributed and decentralized FL-based BGRU model to accurately predict air quality using the smart city's big data. The effectiveness of the FL-based BGRU Model is estimated with other machine learning (ML) models by using various evaluation metrics.