EnerVision

Inspiration

EnerVision: an AI system that blends proven deep learning with cutting-edge architectures to deliver smarter energy forecasting and decision-making. Buildings consume nearly one third of global energy and are a major source for carbon emissions. We were inspired to tackle this challenge by creating a system that not only forecasts energy demand but also learns to make smarter decisions for efficiency and renewable integration.

What it does

EnerVision forecasts short term energy loads using LSTMS & RNNs , detects anomalies and applies DQN to recommend optimal energy saving actions.

Supervised Learning(LSTMs and RNNs) for accurate time series forecasting.
Reinforcement Learning(DQN) for accurate adaptive decision making.
Transformers handle long-term dependencies in energy demand. -** Graph Neural Networks (GNNs)** model interdependencies between multiple buildings and weather factors.

How we built it

Collected and cleaned building energy datasets.
Preprocessed data(normalization, sliding windows, weather).
Built a hybrid forecasting pipeline RNNs/LSTMs for fast baseline forecasting.
Transformer-based models (Informer, PatchTST) for scalable, long-sequence learning.
Integrated GNN layers to capture building-to-building and grid interactions.
Trained a DQN agent for decision-making, extended into multi-agent RL for collaborative optimization.
Evaluated with RMSE, MAE, anomaly detection accuracy, and energy efficiency improvements.

Challenges we ran into

Combining classical deep learning with newer architectures in one pipeline.
Scaling Transformers and GNNs on large datasets while keeping real-time performance.
Designing a reward system that balances cost, carbon footprint, and anomaly mitigation.

Accomplishments that we're proud of

Built a multi-model hybrid AI that outperforms classical ML and standalone LSTMs.
Achieved higher accuracy and smarter decision making than baseline methods.
Demonstrated how forecasting, reinforcement learning and GNNs can transform energy management.

What we learned

RNNs/LSTMs are good for short horizons, but Transformers scale better.
GNNs bring context-awareness by modeling energy networks.
Reinforcement learning (DQN & MARL) turns passive predictions into actionable decisions.

What's next for EnerVision

Real-time deployment in IoT-enabled smart buildings.
Scaling forecasting with Transformer-based Time Series Foundation Models (TSFMs).
Expanding to city-wide smart grids with renewable integration.
Partnering with utilities to reduce energy costs and emissions at scale.