Forecasting greenhouse cucumber production using a combined LSTM-TCN deep neural network.

Kleijkers, Yrsa

Forecasting greenhouse cucumber production using a combined LSTM-TCN deep neural network.

Files

Kleijkers, Y. s-1015940-BSc-Thesis-2022.pdf (1.08 MB)

Authors

Kleijkers, Yrsa

Issue Date

2022-01-30

Language

en

URI

https://theses.ubn.ru.nl/handle/123456789/15873

Abstract

A deep neural network (DNN) is proposed that combines a long short-term memory recurrent neural network (LSTM-RNN) with a temporal convolutional network (TCN) to forecast weekly greenhouse cucumber production. This DNN is expected to overcome the challenges that come along when forecasting greenhouse crop production, like biological unpredictability, behavioral delay in the crop’s response and short and long-term relationships that exist between the crop and the different conditions in the greenhouse. The focus will be on using predictors that can be measured inside the greenhouse to forecast the weekly harvest value, with one or two weeks as input data. Due to limited amount of real-life data, synthetic data is used to optimize and train the LSTM-TCN DNN. This also means that this research mainly focuses on providing a proof of concept rather then creating a ready to use application. The performance of the LSTM-TCN forecasting, expressed by the mean square error (MSE), is compared to using only a LSTM-RNN DNN. The results show an decrease of 69% and 38% in MSE, although not significant, when using the combined LSTM-TCN DNN compared to using a standalone LSTM-RNN DNN for one and two weeks of input data respectively. The results of this Bachelor of Sciences (BSc) thesis research show the potential of using a LSTM-TCN DNN for forecasting greenhouse cucumber production.