pants_brain.js

import brain from 'brainjs';
import _ from 'underscore';
import axios from 'axios';
import Q from 'q';

exports.run = function(state, city, existing_data) {
  if(!existing_data)
    existing_data = [] 
  const deferred = Q.defer();
  let temps = [];
  axios.get(`http://api.wunderground.com/api/22662d2addfdc007/hourly/q/${state}/${city}.json`)
    .then(response => {
      const first_hour = parseInt(response.data.hourly_forecast[0]['FCTTIME']['hour']);
      const yday = parseInt(response.data.hourly_forecast[0]['FCTTIME']['yday']);
      const year = parseInt(response.data.hourly_forecast[0]['FCTTIME']['year']);
      const month = parseInt(response.data.hourly_forecast[0]['FCTTIME']['mon_padded']);
      const day = parseInt(response.data.hourly_forecast[0]['FCTTIME']['mday_padded']);
      const todays = _.filter(response.data.hourly_forecast, function (obj) {
        return parseInt(obj['FCTTIME']['yday']) == yday;
      });
      const evens = _.filter(todays, function (obj) {
        return parseInt(obj['FCTTIME']['hour']) % 2 == 0;
      });
      const needed = _.filter(evens, function (obj) {
        return parseInt(obj['FCTTIME']['hour']) > 5 && parseInt(obj['FCTTIME']['hour']) < 19;
      });
      temps = _.map(needed, (obj) => {
        return {
          hour: "h" + parseInt(obj['FCTTIME']['hour']),
          temp: obj['temp']['english']
        };
      });



      if (temps.length < 7) {
        axios.get(`http://api.wunderground.com/api/22662d2addfdc007/history_${year}${month}${day}/q/${state}/${city}.json`)
          .then(response => {
            const needed = _.filter(response.data.history.observations, function (obj) {
              return ((parseInt(obj['date']['hour']) % 2 == 0) &&
                (parseInt(obj['date']['hour']) > 5) && 
                (parseInt(obj['date']['hour']) < 19)
              );
            });
            const previous_temps = _.map(needed, (obj) => {
              return {
                hour: "h" + parseInt(obj['date']['hour']),
                temp: obj['tempi']
              };
            });
            temps = temps.concat(previous_temps);
            deferred.resolve(computeGuess(temps, existing_data));
          })
          .catch(error => {
            deferred.reject(new Error(error));
          });
      }
      else {
        deferred.resolve(computeGuess(temps, existing_data));
      }
    })
    .catch(error => {
      deferred.reject(new Error(error));
    });
  return deferred.promise;
}



function computeGuess(temps, existing_data) {
  // const net = new brain.NeuralNetwork();

  var net = new brain.NeuralNetwork({
    activation: 'relu' // activation function
  });

  const input = {h6: null, h8: null, h10: null, h12: null, h14: null, h16: null, h18: null}
  _.each(temps, (temp) => {
    const t = Math.round(parseFloat(temp['temp']))
    const v = t < 31 ? 0 : t > 90 ? 1 : (1 * ((t - 30)/60))
    input[temp['hour']] = v
  })


  //patch the missing observations, this weather api is not perfect
  //first try to get the first available previous hour value
  _.each([6,8,10,12,14,16,18], (num) => {
    let i = num
    while(input['h'+num] == null && i > 5){
      input['h'+num] = input['h'+(i-2)]
      i = i-2;
    }
  })
  //first try to get the first available later hour value
  _.each([6,8,10,12,14,16,18], (num) => {
    let i = num
    while(input['h'+num] == null && i < 19){
      input['h'+num] = input['h'+(i+2)]
      i = i+2;
    }
  })


  // below 30 degrees is a 0
  // above 90 degrees is a 1
  // all others fit to scale
  // ex: 70 degrees is 1 * (40/60), it is 40 units into the 60 unit scale
  const training_data = [
    {input: { h6: 0.0, h8: 0.1, h10: 0.2, h12: 0.2, h14: 0.2, h16: 0.2, h18: 0.1 }, output: { pants: 1 }},
    {input: { h6: 0.7, h8: 0.8, h10: 0.9, h12: 1.0, h14: 1.0, h16: 1.0, h18: 0.9 }, output: { shorts: 1 }},
    {input: { h6: 0.5, h8: 0.5, h10: 0.55, h12: 0.6, h14: 0.66, h16: 0.64, h18: 0.6 }, output: { shorts: 1 }},
    {input: { h6: 0.6, h8: 0.65, h10: 0.68, h12: 0.7, h14: 0.73, h16: 0.7, h18: 0.68 }, output: { shorts: 1 }},
    {input: { h6: 0.1, h8: 0.1, h10: 0.3, h12: 0.3, h14: 0.1, h16: 0.1, h18: 0.1 }, output: {pants: 1}},
    {input: { h6: 0.3, h8: 0.33, h10: 0.44, h12: 0.5, h14: 0.44, h16: 0.4, h18: 0.4 }, output: {pants: 1}},
    {input: { h6: 0.35, h8: 0.4, h10: 0.48, h12: 0.52, h14: 0.5, h16: 0.48, h18: 0.47 }, output: {shorts: 1}},
    {input: { h6: 0.03, h8: 0.1, h10: 0.15,h12: 0.2, h14: 0.2, h16: 0.25, h18: 0.2 }, output: {pants: 1}},
    { input: 
      { h6: 0.31666666666666665,
        h8: 0.3,
        h10: 0.3333333333333333,
        h12: 0.4166666666666667,
        h14: 0.4666666666666667,
        h16: 0.5166666666666667,
        h18: 0.5333333333333333 },
     output: { pants: 1 } }
  ]

  net.train(training_data.concat(existing_data))

  const output = net.run(input);  
  return {input, output}
}