Mass flow prediction in a refrigeration machine using artificial neural networks

•An alternative way to estimate the mass flow rate in refrigeration systems using Artificial Neural Networks was developed.•The proposed methodology allowed to estimate the mass flow rate with four inputs using a small Artificial Neural Network.•A low-cost embedded system was proposed to measure all inputs and calculate the mass flow rate. Measuring mass flow rate in refrigeration systems with flow meters can be expensive when taking into account the cost of the equipment itself and the costs related to installation and maintenance. A mode based on Artificial Neural Networks can be used to predict the value of the mass flow, at a low cost, through easily observed and measured parameters, like temperatures. Additionally, well-known correlations to calculate parameters that directly influence the mass flow rate can be used as input data for the model to improve its accuracy. Within this context, the present study aims to develop a Multilayer Perceptrons model to predict the mass flow rate of a refrigeration systems. The model developed was optimized using one hidden layer with four neurons, Adam as the optimizer and softplus as the activation function. Later, it is presented an alternative mass flow rate meter, using an Artificial Neural Network model programmed in a microcontrolled circuit with only three temperatures as inputs. To develop the model, experimental data were collected in a refrigeration machine at several operating points. Step disturbances were introduced in the mass flow rate to produce transient data. Two different data sets were considered in the training process. The first data set contained only steady-state data and in the second data set, there were steady-state plus transient data. The mass flow rate estimated through the proposed model presented an average error of 0.79 % when considering steady-state and transient data in the training process, and 0.81 % when considering only steady-state data in the training procedure. The final embedded system developed predicted the mass flow rate with an average error of 1.22 %.