Abstract:

Forecasting air pollution is tough for smart cities. Populations keep growing, factories keep churning out emissions, and the weather just won’t sit still. Traditional AQI forecasting models, like LSTM networks and other centralized systems, just don’t keep up. They’re too fragile, get tripped up by the complexity of cities, and often miss those sudden, nasty pollution spikes. This paper brings something new to the table: Trans-MARL.

It’s a decentralized multi-agent transformer framework, designed from the ground up for city-wide AQI forecasting. Instead of one big.

fragile model Trans-MARL works as a network a kind of “society of sensors.” Think of it this way: each agent keeps an eye on its own corner of the city they all trade info and together they piece together a clearer more up-to-date picture. The whole thing operates as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP).

Which means every agent only sees a slice of the big picture but by working together they fill in the blanks. Trans-MARL stands on three legs. First, there’s a transformer-based perception layer that catches the time-based patterns.

Then there’s a spatial communication layer, using graph-based embeddings, which is how the agents exchange local info and stay connected. Finally, the multi-agent reinforcement learning decision layer makes sure all the pieces move in the right direction, optimising each agent’s actions for the good of the city. When put to the test, Trans-MARL doesn’t just hold up; it thrives.

It shrugs off sensor failures, snaps into action when pollution patterns shift, and picks up anomalies that older models overlook. In short, this decentralised approach scales with the city. Down the line, it can hook up with satellite images and traffic patterns, pushing environmental monitoring to a smarter, stronger future.

Keywords:

Air Quality Index, Smart Cities, Multi-Agent Reinforcement Learning, Transformer, Networks, Decentralized AI, AQI Forecasting, Urban Intelligence, Environmental Monitoring.

Aim/Objectives

Methods