Adding temperature forecasting model

[meljsingh]

Aim is to add temperature forecasting model onto platform, visualise outputs, and provide relevant warnings.

Explanation of the model

• Inputs: Temperature data for sensors 16B1, 16B2 and 16B4 (table: zensie_data), Energy data for Carpenter's place (table: energy_data)
• Forecasting model: there are two models to compare efficiency (a) static ARIMA model using temperature and lights, (b) dynamic Bayesian model using temperature and a derived component from lights.
• Outputs: 24h mean forecast, upper bound and lower bound, and probability temperatures will surpass threshold

Output table for 1 sensor location:

	Forecast hour	h	h+1	h+2	h+…	h+24
Sensor 16B1	Mean temperature - static
Sensor 16B1	Upper bound - static model
Sensor 16B1	Lower bound - static model
Sensor 16B1	Mean temperature - dynamic
Sensor 16B1	Upper bound - dynamic model
Sensor 16B1	Lower bound - dynamic model

In addition, there are the following four outputs:

• Probability temperature will be over 25degC for over 2 hours
• Number of hours temperatures likely to be over 25degC
• Probability temperature will be under 18degC for over 2 hours
• Number of hours temperatures likely to be under 18degC

Further information:

• a simple crop growth linear regression model can be calculated based on the forecast temperatures to give an estimate of the effect of the conditions on crop health.

Further analysis

The error from the previous day's forecast can provide further information

• Inputs: Previous day's temperature (zensie_data), weather (iweather), energy data (energy_data) & saved forecast from previous day
• Model: calculates the bias error of forecast temperature - observed temperature; calculates change in temperature & lighting, outputs discrepencies
• Outputs:
  • if model error >1degC, provide warning "Temperature varied uncharacteristically yesterday".
  • if average temperature changed by >2deg C, provide explanation "This could be because it was significnalty colder/warmer"
  • if energy was >20% different to average monthly, provide explanation "This could be because you changed your operational pattern".

[meljsingh]

Workflow of forecasting model

in Python

1. Download latest data in R:
2. Clean latest data:
   • take hourly average
   • make sure temperature and energy data is aligned (clock change)
   • construct lights algorithm
   • identify date and time of forecast
   • create training dataset (1 year)
3. Fit model:
   • BSTS dynamic model
   • seasonal ARIMA static model
4. Predict temperatures
   • output mean forecast, upper and lower bounds for each model
   • output probability too hot/ too cold
5. Save outputs in format readable by Python

• [x] Check when stripped down model available for May in OneDrive folder

  This model is in the models folder here

[meljsingh]

To dos (in order of priority

• [x] Mel: Prepare a stripped down version of the forecasting model in R to show May how it works, with all relevant dependencies
• [ ] May: Execute R model, get acquainted and plan how it can be integrated into a python script
• [ ] May & Mel: Write python script which runs R models and saves the data in database
• [ ] May: Create a container in which the models can run at 4 am and 4pm.
• [ ] Mel: Check alignment of temperature and energy data, make necessary adjustments to code to ensure highest possible model accuracy
• [ ] May: Add visualisation of forecast on platform