Artificial neural networks model: reliable forecasting tool in cocoa postharvest losses reduction

Abstract

This research involves the development of an artificial neural network (ANN) model that forecasts the weekly production quantities of outputs for a typical cocoa processing company in order to reduce post-harvest losses. The artificial neural network was initially built with a single input and a single output with the aid of the Neurosolutions 5.07 software package. It was then trained, cross- validated and tested by carrying out a successful pilot test using raw production data obtained from the cocoa processing company. The data set consists of two input variables and two output variables, and the relationship between any input and output variable is complex. Input variables are the weekly quantities of cocoa bags tipped and batches of cocoa nibs roasted, while output variables are weekly quantities of cocoa butter and cocoa cake packaged in cartons. On training the networks, the parameters of specific networks found to give an acceptable mean square error (MSE) were recorded. The network was later modified using different combination types of input(s) and output(s). The model outputs were found to be satisfactory, lying within the defined error limit when compared to the actual outputs. The result shows that the network developed was able to predict the output quantities with a high accuracy, as the training and cross-validation errors at all times both lie within the target error of 0.0001 as specified by the software developers. The network’s ability in forecasting these outputs with a high degree of accuracy goes a long way in demonstrating that artificial neural networks are highly capable of forecasting in situations when there is no closed-formed mathematical relationship between input and output.