Greenhouse4.0-ML

In this work, two 4.0 greenhouses have been analyzed from the data-driven point of view:

• Pleiades: http://ecoproyecta.es/invernadero-bioclimatico/
• Hidroleaf: Experimental greenhouse container developed by several companies in Murcia in conjunction with CEBAS-CESIC and DIIC-FIUM, saves up to 80% of water consumption compared to a traditional greenhouse.

With emphasis on interior temperatures:

• internal air temperature
• inside temperature of the pots

Additionally for the Pleiades greenhouse, a prediction system for the indoor air temperature has been developed, based on automatic statistical models

• auto.arima
• tbats
• snaive
• prophet
• nnetar
• comb (arithmetic mean of the predictions obtained with the previous models)

In order to improve this system, a brief study has been carried out to determine the number of days to be provided to each model in order to obtain a balance between accuracy and consumption of computational resources.

In addition, two scenarios have been developed to make the predictions:

• Un-informed scenario: Only indoor temperature history is available
• Informed scenario: Historicals of the inner temperature are available. In addition, for the following variables, both the historical and the future values are available for 7 days:
  • outside temperature
  • solar radiation
  • external relative humidity

After modifying the cross-validation algorithm proposed by Hyndman & Athanasopoulos, 2018 [Chapter 3.4] the following results were obtained after predicting the indoor temperature at 7 days (horizon forecast = 7, frequency = 30 min):

• Un-informed scenario (crossvalidation ex ante with last 196 days of 2018, 28 weeks)

MODEL	MAPE	RMSE	MAE	CC	DAYS
tbats	9.13	2.95	2.22	15052	15
comb	9.61	3.06	2.31	140553	165
snaive	10.30	3.44	2.53	1	1
prophet	12.11	3.45	2.79	155	165
auto.arima	10.71	3.53	2.70	119777	30
nnetar	14.51	4.63	3.50	4166	106

• Informed scenario (crossvalidation ex-post with last 210 days of 2018, 30 weeks)

MODEL	MAPE	RMSE	MAE	CC	DAYS
comb	6.20	2.06	1.53	281475	150
prophet	6.78	2.12	1.62	368	60
nnetar	6.83	2.16	1.60	12568	60
auto.arima	8.66	3.08	2.24	268538	150
snaive	9.77	3.19	2.38	1	1

Finally, as a practical case, an alert system has been developed, defined as follows

• An alarm is triggered (SNAIVE) if in a week the predicted temperature for broccoli crops exceeds the threshold 40 º C for 6 hours in a row.

	NO ALARM	ALARM
NO ALARM	TNR = 0.82	FPR = 0.17
ALARM	FNR = 0	TPR = 1

ACC = 0.89