System and method for predicting building thermal loads

1. A system for forecasting thermal loads for a building comprising:

a thermal condition forecaster for forecasting weather conditions to be compensated by a building environmental management system; and

a thermal load predictor for modeling environmental management system components and to generate a predicted thermal load for maintaining a set of environmental conditions based at least in part on information representative of the forecasted weather conditions, the thermal load predictor including a recurrent neural network;

wherein a moving median filter smoothes short term data prior to the processing of the recurrent neural network.

2. The system of claim 1 , the recurrent neural network including an input layer having input neurons for receiving building thermal mass data and building occupancy data.

3. The system of claim 2 , the input layer for the recurrent neural network further includes:

an input neuron for receiving local outside air temperature data;

an input neuron for receiving relative humidity data;

an input neuron for receiving solar radiation data;

an input neuron for receiving occupancy data;

an input neuron for receiving data type data; and

an input neuron for receiving measured load data.

4. The system of claim 3 , the input neurons for receiving measured load data further include:

an input neuron for receiving chiller data;

an input neuron for receiving total building cooling load;

an input neuron for receiving thermal storage load; and

an input neuron for receiving total building energy consumption.

5. The system of claim 4 , the input neuron for chiller data further comprising:

an input neuron for chiller supply water temperature;

an input neuron for chiller return water temperature; and

an input neuron for chiller flow rate.

6. The system of claim 4 , the input neuron for thermal storage load further comprising:

an input neuron for supply water temperature data;

an input neuron for return water temperature data; and

an input neuron for flow rate data.

7. The system of claim 1 further comprising:

an automated recurrent neural network optimizer for optimizing recurrent neural network parameters of the recurrent neural network, the automated recurrent neural network optimizer optimizes a selected neural network parameter by minimizing error between values predicted by the neural network and measured values.

8. The system of claim 7 wherein the automated recurrent neural network optimizer selects one of a training set size, a recurrency rate, a learning rate, a momentum, and a number of hidden neurons as a neural network parameter for optimization.

9. The system of claim 7 wherein the automated recurrent neural network optimizer implements a lattice search optimization method to minimize the error between values predicted by the recurrent neural network and measured values.

10. The system of claim 1 further comprising:

a non-weather data function profile generator that maps building thermal load deviations to recurrent neural network input sets so that the thermal load profile generated by the recurrent neural network may be adjusted with the non-weather data function profile.

11. The system of claim 1 , further comprising a memory storing the predicted thermal load.

12. A method for forecasting thermal loads for a building comprising:

collecting building and weather data from a building system and weather forecasting system;

generating a set of forecasted weather conditions and building data from the collected data;

smoothing a short term data set using a moving median filter;

providing the smoothed data set to a first recurrent neural network for the computation of a thermal load prediction; and

operating at least a portion of a building environmental management system based on the thermal load prediction.

13. The method of claim 12 further comprising:

archiving the data set; and

generating a training data set from the archived data.

14. The method of claim 13 further comprising:

comparing the thermal load prediction generated by the first recurrent neural network to actual measurements taken at the prediction time; and

determining whether a fault detection has occurred.

15. The method of claim 14 further comprising:

archiving the actual measurements; and

generating training predictions for the second recurrent neural network.

16. The method of claim 15 further comprising:

comparing the generated thermal predictions from the second recurrent neural network to the training predictions; and

adjusting neuron weights for neurons in the second recurrent neural network.

17. The method of claim 16 further comprising:

replacing the first recurrent neural network with a copy of the second recurrent neural network in response to a fault detection.

18. A method for forecasting thermal loads for a building comprising:

forecasting weather conditions to be compensated by a building environmental management system; and smoothing a short term data set using a moving median filter; generating a predicted thermal load for maintaining a set of environmental conditions using a recurrent neural network based at least in part on the smoothed information representative of the forecasted weather conditions;

generating a predicted thermal load for maintaining a set of environmental conditions using a neural network based at least in part on the information representative of the forecasted weather conditions; and

operating at least a portion of the building environmental management system based on the thermal load prediction.

19. The method of claim 18 further comprising:

optimizing recurrent neural network parameters of the recurrent neural network by minimizing error between values predicted by the neural network and measured values.

20. The method of claim 19 wherein one of a training set size, a recurrency rate, a learning rate, a momentum, and a number of hidden neurons for the recurrent neural network is optimized.

21. The method of claim 19 wherein the recurrent neural network parameter optimization is performed using a lattice search optimization method.

22. The method of claim 19 further comprising:

generating a non-weather data function profile that maps building thermal load deviations to recurrent neural network input sets; and

adjusting the thermal load profile generated by the recurrent neural network with the non-weather data function profile.