In energy management, we are obsessed with prediction accuracy. We train neural networks to minimize statistical error metrics (e.g., MSE). But here is the hard truth we found in our latest research: A more accurate model does not always lead to cheaper operations. Standard methods optimize for the wrong metric. They minimize prediction error, not operational cost.
In our new paper accepted in IEEE Transactions on Smart Grid, we propose a Decision-Focused Learning (DFL) approach using stochastic smoothing to learn the thermal dynamics of buildings.
🚀 The result? We train the NN to directly minimize the bill, not the error.
We demonstrate that this approach outperforms standard methods for the HVAC scheduling of a realistic 5-zone building in Denver, utilizing:
✅ Mixed-Integer Quadratic Programming (MIQP)
✅ Reformulation of the Neural Network as optimization constraints with adaptive Big-M bounds
✅ Stochastic smoothing to differentiate through the discrete optimization
If you are working on decision-making tool involving machine learning models, this is the shift we need to make.
📄 Read the full paper here.