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

This paper proposes a bi-directional feedback fusion framework that integrates human activity embeddings with dual-timescale temporal modules to significantly improve the accuracy and interpretability of indoor CO2 and PM2.5 forecasting compared to traditional data-driven models.

The Gist

Imagine you are trying to predict the weather inside your house. But instead of just looking at the thermometer and the humidity gauge, you also know exactly what everyone in the house is doing: cooking dinner, opening a window, or sleeping.

This paper presents a new "super-forecasting" system for indoor air quality. It's designed to predict two invisible but important things: CO2 (the gas we breathe out) and PM2.5 (tiny dust particles from cooking or cleaning).

Here is the simple breakdown of how it works, using some everyday analogies.

The Problem: The "Blind" Forecaster

Traditional air quality models are like a blind meteorologist. They only look at the history of the temperature and air sensors.

• The Flaw: If someone suddenly starts frying bacon, the air fills with smoke (PM2.5) instantly. A blind model only sees the temperature rising slowly and thinks, "Oh, it's just a warm day." It misses the sudden spike because it doesn't know someone is cooking.
• The Result: These old models are great at predicting slow changes (like CO2 building up while people sleep) but terrible at predicting sudden, dangerous spikes caused by human actions.

The Solution: The "Two-Stream" Detective

The authors built a new AI that acts like a detective with two eyes. It doesn't just look at the air; it watches the people, too.

1. The Two Streams (The Eyes)

The system processes information in two parallel lanes:

• Lane A (The Environment): This reads the sensors (temperature, humidity, current air levels). It's like looking at the thermometer.
• Lane B (The Action): This reads a log of what people are doing (e.g., "cooking," "cleaning," "opening a window"). It's like watching a security camera of human behavior.

2. The "Bi-Directional Feedback" (The Conversation)

This is the magic sauce. In old models, you might just smash the two data streams together (like putting a thermometer and a camera in a blender). That doesn't work well.

Instead, this new system lets the two streams talk to each other in a loop.

• The Analogy: Imagine a Chef (Environment) and a Waiter (Action) in a busy restaurant.
  • The Waiter sees a customer order a spicy dish (Action). He whispers to the Chef, "Get ready, smoke is coming!"
  • The Chef looks at the stove and says, "I see the smoke rising, but I also know the ventilation is slow."
  • They keep passing notes back and forth. The Waiter adjusts his prediction based on the Chef's view of the stove, and the Chef adjusts his cooking based on the Waiter's news.
• In the AI: The "Action" stream tells the "Environment" stream, "Hey, someone is frying, expect a smoke spike!" The "Environment" stream tells the "Action" stream, "I see the air is already thick, so that spike will last longer." They refine their guess together, over and over again, until they get it right.

3. The "Dual Timescale" (The Long and Short Memory)

The system has two different types of memory to handle different speeds of change:

• The Long-Term Memory (The Slow Turtle): This handles CO2. CO2 builds up slowly, like water filling a bathtub. It takes hours to notice a change. This part of the AI is patient and looks at the big picture.
• The Short-Term Memory (The Fast Rabbit): This handles PM2.5. Smoke from frying appears in seconds and disappears quickly. This part of the AI is hyper-alert and reacts instantly to sudden events.
• Why it matters: If you use a slow turtle to predict a fast rabbit, you miss the race. If you use a fast rabbit to predict a slow turtle, you get jittery and confused. This AI uses both, so it's perfect for both.

4. The "Uncertainty" (The Confidence Meter)

Finally, the AI doesn't just give a number; it gives a confidence score.

• The Analogy: If you ask a human, "Will it rain?" they might say, "Yes, 100% sure" (if the sky is black) or "Maybe, 50/50" (if it's cloudy).
• The AI: If the air is chaotic (like during a party with cooking and dancing), the AI says, "I predict high pollution, but I'm only 60% sure because things are changing fast." This helps building managers know when to trust the prediction and when to double-check manually.

The Results

The team tested this on real data from homes and offices.

• Old Models: Missed the cooking spikes and were often wrong about sudden pollution.
• New Model: Because it "knew" people were cooking, it predicted the smoke spikes accurately. It also predicted the slow CO2 buildup perfectly.
• The Verdict: By combining "what the air feels like" with "what people are doing," the system became much smarter, safer, and more reliable.

In a Nutshell

This paper teaches us that to predict indoor air quality, you can't just look at the air. You have to understand the people inside. By creating an AI that lets the "air sensors" and the "human activity logs" have a conversation, we can finally predict when the air will get bad, giving us time to open a window or turn on a fan before it becomes a health risk.

Technical Summary

Here is a detailed technical summary of the paper "Bi-Directional Feedback Fusion for Activity-Aware Forecasting of Indoor CO2 and PM2.5".

1. Problem Statement

Indoor Air Quality (IAQ) forecasting is critical for occupant health and smart building control. However, predicting future concentrations of key pollutants like Carbon Dioxide (CO2) and Fine Particulate Matter (PM2.5) is challenging due to:

• Complex Interplay: The interaction between environmental factors (ventilation, temperature) and highly dynamic human behaviors (cooking, cleaning, occupancy changes).
• Limitations of Traditional Models: Existing data-driven models rely heavily on historical sensor trajectories. They often fail to anticipate behavior-induced emission spikes (e.g., sudden PM2.5 spikes from frying) or rapid concentration shifts because they lack explicit modeling of occupant actions.
• Temporal Mismatch: CO2 accumulation follows slow, long-term trends driven by occupancy, while PM2.5 fluctuations are often short-term, event-driven, and irregular. A single temporal modeling approach struggles to capture both.

2. Methodology

The authors propose a Dual-Stream Bi-Directional Feedback Fusion Framework that jointly models environmental evolution and human activity.

A. Input Streams

The framework processes two complementary data streams:

1. Environmental Stream ($X_e$): Time-series sensor data (temperature, humidity, CO2, PM2.5).
2. Action Stream ($X_a$): Human activity logs converted into embeddings. Activity labels (e.g., "frying," "window opening") are encoded using the all-MiniLM-L6-v2 model to create semantic vector representations.

B. Model Architecture

The architecture consists of five key components:

1. Individual Stream Encoders:

   • Each stream is processed independently by a Stream-wise Temporal Encoder (STE) to capture unique temporal dynamics (environmental trends vs. behavioral patterns) before fusion.
   • The action stream embeddings are processed to capture temporal semantic cues.

2. Shared and Private Representations:

   • The encoded features are linearly transformed into two subspaces:
     • Shared ($S$): Captures common latent attributes (e.g., general occupancy effects).
     • Private ($P$): Captures stream-specific cues (e.g., cooking spikes specific to PM2.5).
   • This separation prevents one modality from dominating the other.

3. Bi-Directional Feedback Fusion (Core Innovation):

   • Instead of simple concatenation, the model uses an iterative feedback loop ($R$ rounds).
   • Context-Aware Modulation: A global fusion state ($F$) generates modulation parameters ($\gamma, \beta$) that scale and shift the private representations of each stream. This allows the model to adaptively emphasize behavioral cues during transient events or environmental trends during stable periods.
   • Iterative Refinement: The modulated streams feed updated information back into the global state, allowing the model to jointly reason over environmental and behavioral dynamics.

4. Dual-Timescale Temporal Modules:

   • Long-term Module (GRU): Operates on the fused state to capture slow-varying patterns (e.g., CO2 accumulation, ventilation cycles).
   • Short-term Module (GRU): Operates on the concatenation of private subspaces to capture rapid, high-frequency variations (e.g., cooking spikes).
   • The outputs are concatenated to form a comprehensive temporal representation.

5. Prediction Head:

   • Maps the temporal representation to future pollutant levels ($\hat{Y}$) for CO2 and PM2.5.

C. Training Objectives

The model is trained using a composite loss function:

• Primary Loss: Weighted Mean Squared Error (MSE) combined with Negative Log-Likelihood (NLL) for uncertainty estimation.
• Uncertainty Modeling: The model predicts both the mean ($\mu$) and log-variance ($\log \sigma$). The study compares Homoscedastic (constant variance) and Heteroscedastic (input-dependent variance) uncertainty, finding the latter superior for volatile conditions.
• Regularization Terms:
  • Alignment ($R_{align}$): Ensures shared representations point in similar directions.
  • Independence ($R_{ind}$): Penalizes correlation between global and private features to ensure distinct information.
  • Diversity ($R_{div}$): Encourages private features from different streams to be non-redundant.

3. Key Contributions

1. Dual-Stream Framework: A novel architecture that explicitly integrates environmental sensor data with language-derived action embeddings for IAQ forecasting.
2. Bi-Directional Feedback Fusion: Introduces a context-aware modulation mechanism that iteratively refines cross-stream interactions, outperforming static concatenation or attention-based fusion.
3. Dual-Timescale Modeling: Specialized modules that simultaneously capture long-term CO2 trends and short-term PM2.5 fluctuations.
4. Robust Loss Formulation: A composite objective integrating spike-aware penalties and heteroscedastic uncertainty regularization, enhancing performance under volatile indoor conditions.
5. Interpretability: The model provides uncertainty estimates, crucial for decision-support systems in smart buildings.

4. Experimental Results

The framework was evaluated on the DALTON dataset (30 indoor environments, 6 months of data).

• Performance Metrics: The proposed model significantly outperformed state-of-the-art baselines (iTransformer, Multi-Stage TCN, Patch-TST) and ablation variants.
  • CO2: Achieved $R^2 = 0.93$ and RMSE = 81.47.
  • PM2.5: Achieved $R^2 = 0.88$ and RMSE = 30.43.
• Ablation Studies:
  • Fusion Strategy: Bi-directional feedback with 3 iterations ($R=3$) yielded the best results. Removing feedback or using simple concatenation caused significant performance drops (e.g., PM2.5 $R^2$ dropped from 0.88 to 0.37 with direct concatenation).
  • Temporal Modeling: The custom STE encoder outperformed standard Transformers and TCNs, particularly in handling the mixed dynamics of CO2 and PM2.5.
  • Loss Function: Heteroscedastic uncertainty loss improved PM2.5 RMSE by 36% compared to MSE-only, demonstrating the value of adaptive uncertainty modeling.
  • Dual-Timescale: Using both short and long-term modules improved RMSE by ~18% for CO2 and ~20% for PM2.5 compared to single-scale models.
• Qualitative Analysis: The model successfully tracked general trends and event timings (e.g., cooking, bedroom occupancy) but, as an open-loop autoregressive model, showed slight delays and underestimation during sharp, sudden peaks.

5. Significance and Conclusion

This work addresses a critical gap in IAQ forecasting by treating human behavior as a distinct, actionable input stream rather than just a latent variable.

• Practical Impact: The ability to predict behavior-induced spikes (like PM2.5 from cooking) allows for proactive ventilation control, improving energy efficiency and health outcomes.
• Robustness: The integration of uncertainty estimation makes the model suitable for real-world deployment where sensor noise and missing data are common.
• Future Directions: The authors suggest extending the framework to closed-loop forecasting (using real-time sensor corrections) and incorporating broader behavioral signals (wearables, vision).

In summary, the paper demonstrates that integrating behavioral context with environmental sensing via bi-directional feedback fundamentally reshapes model expressiveness, leading to state-of-the-art accuracy and reliability in predicting complex indoor air quality dynamics.