Table 4 Comparative analysis of AiM and baseline models with accuracy and efficiency metrics.
From: AiM: urban air quality forecasting with grid-embedded recurrent MLP model
Model | Avg. Accuracy | Avg. Loss | Latency (ms) | Params (k) | Strengths | Limitations |
|---|---|---|---|---|---|---|
ARIMA26 | 60% | 0.42 | 0.5 | 5 | Well-established statistical forecasting | Struggles with non-stationarity and multi-dimensional dependencies |
SARIMAX27 | 63% | 0.40 | 0.7 | 6 | Incorporates exogenous variables | Limited scalability to high-frequency AQI fluctuations |
SVR28 | 70% | 0.35 | 2.3 | 20 | Captures non-linear trends | No sequential modeling; poor in long-horizon forecasts |
RF29 | 78% | 0.28 | 1.8 | 50 | Handles feature importance well | Weak in temporal dependency learning |
XGBoost30 | 82% | 0.22 | 2.1 | 45 | Excellent tabular performance | Requires heavy tuning; lacks temporal dynamics |
GRU53 | 84% | 0.18 | 5.2 | 120 | Efficient sequence modeling | Loses detail in long-term dependencies |
LSTM52 | 85% | 0.16 | 5.8 | 130 | Captures long-term dependencies | Gradient vanishing/exploding under extreme fluctuations |
CNN-LSTM10 | 86% | 0.14 | 6.3 | 150 | Extracts spatial + temporal features | Higher complexity; prone to overfitting on small datasets |
BLSTM54 | 88% | 0.12 | 8.5 | 220 | Bi-directional learning of temporal features | Computationally expensive; still noise-sensitive |
FL-BGRU11 | 89% | 0.11 | 9.1 | 230 | Federated + bi-GRU improves generalization | Dependency on distributed data quality |
TinyML GRU (Edge AI)55 | 82% | 0.22 | 3.1 | 60 | Ultra-low latency; edge deployable | Accuracy trade-off due to quantization |
Proposed AiM | 96% | 0.04 | 4.5 | 125 | Grid-embedded Bi-GRU captures spatiotemporal patterns; robust to noise | Slightly higher initial training overhead |
