SmartMeterPrediction - HydroQuebec

Home energy usages prediction based on machine learning and smart meter data

Getting Started

• Go to the docker folder
• Build the docker containers in the sub-folder
• In docker-compose.yml add the necessary informations
• docker-compose up
• Graphana setup
  • go to localhost:3000
  • add the data source
  • url : http://influxdb:8086
  • Use Proxy setting
  • Database name : e
  • import energy_usage.json as a new dashboard
• A new prediction should be made everyday

Training

Pre-reqs

• pip install -r requirements.txt in the training folder
• tensorflow-gpu is strongly recommended if you have the hardware

Multiple linear regression

• Download your house's hourly datasets from Hydro-Quebec
• Put them all in the the ./training/data/hourly folder
• Run the cvs_hour_processing.py script. This should create a new csv file. Make sure that there are no hours with no energy usage in the data set
• Go to the multi_linear_regression_hourly.py and change line 20 to point the new csv created
• Run multi_linear_regression_hourly.py
• Uses the temperature, hour of the day and is_workday to make predictions and train the network
• 

LSTM network

• Download your house's hourly datasets from Hydro-Quebec
• Put them all in the the ./training/data/hourly folder
• Run the cvs_hour_processing.py script
• Make sure that there are no hours with no energy usage in the data set
• Go to the lstm_hourly.py and change line 43 to point the new csv created
• You can adjust what parameters we want to use for training by changing line 36 and the numbers of prediction hours in line 39 - line 41
• Uses the temperature, hour of the day previous predictions and is_workday to generate new predictions and train the network
• NOTE: I've currently commented out in the LSTM prediction because it's not very accurate. To re-enable you need to remove the comments in ./predictor/main.py and ./predictor/database.py
•