(596h) Using Residual Neural Networking and Lagrange Optimization to Predict Cooling Demand and Calculate Optimal Chiller Settings

Deep learning (i.e. neural networking) methods has become more prevalent across many different subfields in engineering. Specifically, these deep learning methods can improve operational efficiencies by predicting user demand. For warmer climates, cooling demand prediction has a large potential for savings, so we piloted a deep learning based optimization system on the University of Arizona central cooling loop.

We propose a two module system which predicts cooling demand and calculates optimally efficient chiller settings which can be adapted to most multi-chiller central cooling systems. To predict cooling demand for a serviceable area, a neural network was used with historical load, weather, and other user indicators. Convolutional (1-dimensional) and gated recurrent neural architectures work well with time sequenced features such as past load and weather [1]. Novel gains in prediction accuracy are obtained by adding an input layer of the most recent past demand to the final layer of the network. In effect, this residual like layering of past load compels the network to predict the demand difference from the previous day. Switching to a difference-based output decreased the mean absolute error of energy demand by 40%. To find the optimal settings for a set of connected chillers, LaGrange optimization was used [2]. With this system, operators can set optimal chiller settings for variable lengths of time.

The proposed system was implemented on a central cooling loop that supports the hospital, research laboratories, classrooms, sports facilities, and offices with up to 27 MW of cooling. Past load, weather, class scheduling, sports scheduling, and date indices are used as features to train and evaluate the offset convolutional and recurrent neural network. The data was split into testing and training data to allow model evaluation without data leakage or overfitting. On an AWS p2.xlarge instance with GPU enabled Tensorflow, training for 10 epochs took around 50 minutes. On a CPU-based computer (4 core), the model can make predictions in about 7-10 minutes. The load prediction model and the LaGrange optimization computed optimal chiller flowrates at hourly intervals. The overall system was evaluated based on system energy use and overall cost savings. Based on historical testing data, cost savings were estimated to be in the tens of thousands per year.

[1] He, W. (2017). Load forecasting via deep neural networks. Procedia Computer Science, 122, 308-314.

[2] Chang, Y. C. (2004). A novel energy conservation method—optimal chiller loading. Electric Power Systems Research, 69(2-3), 221-226.