Bi Directional Feedback Fusion for Activity Aware Forecasting of Indoor CO2 and PM2.5

Harshala Gammulle; Lidia Morawska; Sridha Sridharan; Clinton Fookes

cs.LGMar 6, 2026

Bi Directional Feedback Fusion for Activity Aware Forecasting of Indoor CO2 and PM2.5

Authors: Harshala Gammulle, Lidia Morawska, Sridha Sridharan, Clinton Fookes

Organizations: Signal Processing, Artificial Intelligence and Vision Technologies (SAIVT) Lab, Queensland University of Technology, Brisbane, Australia. · International Laboratory for Air Quality and Health (ILAQH), Queensland University of Technology, Brisbane, Australia.

Abstract

Indoor air quality (IAQ) forecasting plays a critical role in safeguarding occupant health, ensuring thermal comfort, and supporting intelligent building control. However, predicting future concentrations of key pollutants such as carbon dioxide (CO2) and fine particulate matter (PM2.5) remains challenging due to the complex interplay between environmental factors and highly dynamic occupant behaviours. Traditional data driven models primarily rely on historical sensor trajectories and often fail to anticipate behaviour induced emission spikes or rapid concentration shifts. To address these limitations, we present a dual stream bi directional feedback fusion framework that jointly models indoor environmental evolution and action derived embeddings representing human activities. The proposed architecture integrates a context aware modulation mechanism that adaptively scales and shifts each stream based on a shared, evolving fusion state, enabling the model to selectively emphasise behavioural cues or long term environmental trends. Furthermore, we introduce dual timescale temporal modules that independently capture gradual CO2 accumulation patterns and short term PM2.5 fluctuations. A composite loss function combining weighted mean squared error, spike aware penalties, and uncertainty regularisation facilitates robust learning under volatile indoor conditions. Extensive validation on real-world IAQ datasets demonstrates that our approach significantly outperforms state of the art forecasting baselines while providing interpretable uncertainty estimates essential for practical deployment in smart buildings and health-aware monitoring systems.

Figures & tables

Fig. 1 : Overview of the proposed dual‑stream bi-directional feedback fusion framework for indoor air‑quality forecasting. The model processes environmental sensor data and action embeddings through separate stream encoders to capture environmental evolution and human activity patterns. A bi-directional feedback fusion module iteratively aligns and refines cross‑stream representations using shared‑private decomposition and context‑aware modulation. Dual‑timescale modules then model both long‑term CO 2 trends and short‑term PM 2.5 fluctuations before generating pollutant-specific predictions.

$L_{l}$	$L_{p}$	CO 2			PM 2.5
$L_{l}$	$L_{p}$	RMSE $\downarrow$	MAE $\downarrow$	R 2 $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$	R 2 $\uparrow$
24	5	99.75	68.43	0.89	39.42	29.40	0.80
36	10	78.14	40.43	0.94	32.63	24.23	0.86
48	15	81.47	53.15	0.93	30.43	24.06	0.88
60	15	76.09	50.97	0.94	35.42	30.42	0.84
60	30	148.11	82.01	0.79	52.32	44.01	0.65

TABLE I : Experimental results across different forecasting horizons. Here,

L_{l}

and

L_{p}

denote the lookback and prediction windows, respectively.

Model	CO 2			PM 2.5
Model	RMSE $\downarrow$	MAE $\downarrow$	R 2 $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$	R 2 $\uparrow$
$X_{e}$ only	134.16	71.27	0.81	65.60	43.75	0.44
$X_{e}+X_{a}$ (direct concat)	147.06	86.38	0.77	70.04	50.29	0.37
$X_{e}+X_{a}$ (2-streams; concat)	125.82	62.59	0.83	59.92	39.63	0.54
Ours (no feedback, $R{=}0$ )	123.37	65.30	0.84	39.10	19.56	0.80
Ours (feedback, $R{=}3$ )	81.47	53.15	0.93	30.43	24.06	0.88

TABLE II : Impact of dual-stream information and fusion strategy on IAQ forecasting performance.

Model	CO 2			PM 2.5
Model	RMSE $\downarrow$	MAE $\downarrow$	R 2 $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$	R 2 $\uparrow$
iTransformer	128.23	84.68	0.82	49.96	23.52	0.68
Multi-Stage TCN (3 stages)	120.68	73.02	0.84	41.16	33.04	0.78
patch-TST	100.03	61.24	0.89	35.04	14.02	0.84
STE(Ours)	81.47	53.15	0.93	30.43	24.06	0.88

TABLE III : Impact of different temporal modelling strategies on IAQ forecasting performance.

Fusion Variant	CO 2			PM 2.5
Fusion Variant	RMSE $\downarrow$	MAE $\downarrow$	R 2 $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$	R 2 $\uparrow$
Direct Concat	147.06	86.38	0.77	70.04	50.29	0.37
Two-Streams Concat	125.82	62.59	0.83	59.92	39.63	0.54
No feedback ( $R{=}0$ )	123.37	65.30	0.84	39.10	19.56	0.80
Feedback $R{=}1$	92.42	47.40	0.91	35.04	14.02	0.84
Feedback $R{=}3$	81.47	53.15	0.93	30.43	24.06	0.88

TABLE IV : Impact of bi-directional feedback fusion on IAQ forecasting performance.

Temporal Variant	CO 2			PM 2.5
Temporal Variant	RMSE $\downarrow$	MAE $\downarrow$	R 2 $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$	R 2 $\uparrow$
Short-term only	100.39	74.22	0.90	36.65	28.23	0.83
Long-term only	99.15	71.23	0.90	38.39	29.20	0.81
With dual-timescale	81.47	53.15	0.93	30.43	24.06	0.88

TABLE V : Impact of dual-timescale temporal modules on forecasting performance.

Loss	CO 2			PM 2.5
Loss	RMSE $\downarrow$	MAE $\downarrow$	R 2 $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$	R 2 $\uparrow$
$\mathcal{L}_{\mathrm{MSE}}$ only	93.61	43.05	0.91	47.88	35.69	0.70
$\mathcal{L}_{\mathrm{MSE}}+\mathcal{L}_{\mathrm{NLL}}$ (Homo)	87.83	51.53	0.92	45.13	37.39	0.74
$\mathcal{L}_{\mathrm{MSE}}+\mathcal{L}_{\mathrm{NLL}}$ (Hetero)	81.47	53.15	0.93	30.43	24.06	0.88

TABLE VI : Impact of loss formulation on IAQ forecasting performance.

Regularisation	CO 2			PM 2.5
Regularisation	RMSE $\downarrow$	MAE $\downarrow$	R 2 $\uparrow$	RMSE $\downarrow$	MAE $\downarrow$	R 2 $\uparrow$
$\mathcal{L}_{\star}$ only	105.11	55.49	0.89	52.10	30.19	0.65
$\mathcal{L}_{\star}$ + $\mathcal{R}_{\text{align}}$	101.09	66.23	0.89	37.75	30.59	0.82
$\mathcal{L}_{\star}$ + $\mathcal{R}_{\text{align}}+\mathcal{R}_{\text{ind}}$	83.12	38.12	0.92	36.82	26.35	0.82
$\mathcal{L}_{\star}$ + $\mathcal{R}_{\text{align}}+\mathcal{R}_{\text{ind}}+\mathcal{R}_{\text{div}}$	81.47	53.15	0.93	30.43	24.06	0.88

TABLE VII : Impact of regularisation terms on IAQ forecasting performance

Air Quality Prediction

Air quality prediction estimates future pollutant levels and maps local exposure to support health warnings, planning, and pollution control.

See all 43 papers

AirFlow: Context Preserving and Multi-Rate State Modeling for Air Quality Forecasting
Aug 10, 2026Fan Yang, Nan Chen, Yijie Dong +2Multivariate Time Series Forecasting Air Quality Prediction
School of Software Technology, Zhejiang University Hangzhou, Zhejiang 310000, China
Interpretable PM2.5 Forecasting for Urban Air Quality: A Comparative Study of Adaptive Time-Series Models
Mar 26, 2026Moazzam Umer Gondal, Hamad ul Qudous, Asma Ahmad Farhan +1Air Quality Prediction Time Series Forecasting
School of Computing, National University of Computer & Emerging Sciences (FAST), Lahore 54000, Pakistan · College of Computing and Informatics, Saudi Electronic University, Riyadh, 11673, Saudi Arabia
AirMPA: A Meteorology-to-Pollution Adapter for Global Air Quality Forecasting
Oct 19, 2026Yiheng Wang, Kai Zheng, Yuetan Lin +5Air Quality Prediction Spatiotemporal Forecasting
Fudan University · Shanghai Academy of AI for Science · University of Shanghai for Science and Technology

Aug 10, 2026J/K move · Enter open · S save

AirFlow: Context Preserving and Multi-Rate State Modeling for Air Quality Forecasting

Fan Yang, Nan Chen, Yijie Dong, Yuchen Zhang+1

School of Software Technology, Zhejiang University Hangzhou, Zhejiang 310000, China

Multivariate Time Series Forecasting Air Quality Prediction Non-Stationary Time Series Forecasting Multiscale Time-Series Modeling State Space Models

Open paper AskPDF arXiv

Accurate air quality forecasting is essential for public health and urban environmental management, but remains challenging because pollutant channels differ in periodicity and distribution drift, while their concentration trajectories contain both multi-scale dependencies and rapid changes. Recent methods have improved spatial dependency learning and meteorological covariate modeling. However, pollutant channels are still passed through the same normalization rule and temporal backbone, using a shared latent representation for channel-specific distributions and changes at different rates. To address this limitation, we propose AirFlow, a pollutant-aware dual-stream framework that operates on station multivariate observations without additional graph propagation or predefined signal decomposition. Specifically, AirFlow designs two novel blocks: (1) a statistic-guided normalization routing mechanism that selects a normalization path for each pollutant according to its 24-hour autocorrelation and distribution drift; and (2) a hierarchical dual-stream state model that combines multi-scale state space propagation with learnable response coefficients, where gated bidirectional cross-attention exchanges information and adaptively fuses the resulting representations. Experiments on real-world data from multiple cities show that AirFlow achieves the best performance in 34 of 36 metrics comparisons, with reductions of up to 11.11% root mean square error over the state-of-the-art baseline. AirFlow also requires only 0.0483M parameters and 0.0215G FLOPs, achieving high forecasting accuracy with low computational overhead.

Abstract

Air Quality Prediction

Explore similar work

AirFlow: Context Preserving and Multi-Rate State Modeling for Air Quality Forecasting

Interpretable PM2.5 Forecasting for Urban Air Quality: A Comparative Study of Adaptive Time-Series Models

AirMPA: A Meteorology-to-Pollution Adapter for Global Air Quality Forecasting