Near-instant g-function construction with artificial neural networks
Abstract
A g-function is a useful tool that simplifies the calculations of heat exchanges in a ground-coupled heat pump system. In this work, we show how an artificial neural network can be trained to construct a g-function with high efficiency and reliability. First, we show how a block matrix formulation can be used to construct rapidly a g-function. This method is then used to assemble a database of 27,000 g-functions with a variety of input parameters. This database of g-functions is used to train a feed-forward neural network having three hidden layers using the back-propagation algorithm to update the weights and biases of the neurons. The network we developed in this work can estimate the long-term g-function of a ground heat exchanger made of 1 to 10 boreholes over a duration of 100 years with various ground thermal properties, borehole field configurations, length and buried depth in a few milliseconds. The contribution of this work is to lay out the methodology to allow anyone to construct a g-function with an artificial neural network.