PresentationPDF Available

Using Artificial Neural Networks to Predict Climate in a Greenhouse

Authors:

Abstract

We present an application of time-series prediction to the climate inside a greenhouse. We use an Artificial Neural Network (ANN) model to predict the air temperature and relative humidity inside the greenhouse 5 minutes in advance. Such a model generates a One Step Prediction (OSP). We iterate this OSP to get Long-Term Predictions (LTP), thus generating a complete forecasting series from a single starting point. A minimum set of input signals included the following: Air temperature and relative humidity, ventilation opening, thermal screen opening, and 4 astronomical parameters (hour angle, de-clination, elevation and theoretical radiation). All 8 input variables were also fed with a 3-steps delay (t={0,-1,-2}) to predict the next step (t=+1). For training the model, we used data from three complete production periods, 2010 through 2013 in two tomato greenhouses in Berlin-Dahlem. Data was measured every 5 minutes, filtered , randomized and separated into a train, validation and a test datasets, with 545,156, 96,204 and 119,804 records, respectively. The results show that the error can be as low as 0.5 • C for temperature and 2% relative humidity , when predicting a single step ahead. However, the use of these predicted values to feed the model recursively leads to an increment in the uncertainty, thus limiting the LTP to a maximum of three or four steps. We show different cases where this prediction seems to follow the actual trend in the temperature and relative humidity signals, as well as some where it shoots up and the error increases rapidily, making the prediction unfit for control or decision making. These examples give us material and hints on how to improve our predictive greenhouse models, aiming at a predictive, non-reactive control system.
VORTRÄGE / SESSION 3 / PROCESS CONTROL
Using Artificial Neural Networks to Predict Climate in a Greenhouse
Luis Miranda, Dennis Dannehl, Ingo Schuch, Thorsten Rocksch, Uwe Schmidt
Humboldt-Universität zu Berlin
luis.carlos.miranda.trujillo@cms.hu-berlin.de
We present an application of time-series prediction to the climate inside a greenhouse. We use
an Artificial Neural Network (ANN) model to predict the air temperature and relative humidity
inside the greenhouse 5 minutes in advance. Such a model generates a One Step Prediction
(OSP). We iterate this OSP to get Long-Term Predictions (LTP), thus generating a complete
forecasting series from a single starting point.
A minimum set of input signals included the following: Air temperature and relative humidity,
ventilation opening, thermal screen opening, and 4 astronomical parameters (hour angle, de-
clination, elevation and theoretical radiation). All 8 input variables were also fed with a 3-steps
delay (t={0,-1,-2}) to predict the next step (t=+1).
For training the model, we used data from three complete production periods, 2010 through
2013 in two tomato greenhouses in Berlin-Dahlem. Data was measured every 5 minutes, filte-
red, randomized and separated into a train, validation and a test datasets, with 545,156, 96,204
and 119,804 records, respectively.
The results show that the error can be as low as 0.5 C for temperature and 2% relative humidi-
ty, when predicting a single step ahead. However, the use of these predicted values to feed the
model recursively leads to an increment in the uncertainty, thus limiting the LTP to a maximum
of three or four steps. We show different cases where this prediction seems to follow the actual
trend in the temperature and relative humidity signals, as well as some where it shoots up and
the error increases rapidily, making the prediction unfit for control or decision making. These ex-
amples give us material and hints on how to improve our predictive greenhouse models, aiming
at a predictive, non-reactive control system.
12 BHGL-Tagungsband 31/2015
ResearchGate has not been able to resolve any citations for this publication.
ResearchGate has not been able to resolve any references for this publication.