public void trainingEnsemble(List<timeSeries> tslocal, forecastSettings _fs, bool additionalTraining)
{
trainingSet = new List<StructuredDataSet>();
for (int ts = 0; ts < n_timestamps; ts++)
{
//create training set
//Console.WriteLine("train nns for timestamp = " + ts.ToString() + " out of 95");
trainingSet.Add(new StructuredDataSet());
cvSet.Add(new StructuredDataSet());
double valueMax = 0;
double valueMin = double.MaxValue;
for (int i = 0; i < tslocal[ts].timeseries.Count; i++)
{
//if (tslocal[ts].timeseries[i] == 0)
{
var value = tslocal[ts].timeseries[i];
if (value > valueMax) valueMax = value;
if (value < valueMin) valueMin = value;
}
}
maxValue[ts] = valueMax;
minValue[ts] = 0.0;// valueMin;//_ts.timeseries.Average() / maxValue; //
maxValueT[ts] = tslocal[ts].temperature.Max();
meanValueT[ts] = 0.0;//_ts.timeseries.Average() / maxValue; //
trainingSet[ts].Pairs = new List<DataPair>();
//преобразуем то, что нужно в выборки данных.
//формируем trainingset
for (int time = _fs.energyLags.Max(); time < tslocal[ts].timeseries.Count(); time++)
//var time = tslocal[ts].timeseries.Count - 1;
{
//if (tslocal[ts].outlierLabel[time] == 0)
{
DataPair trainingPair = new DataPair();
trainingPair.InputVector = new List<double>();
trainingPair.InputVector = createInputVector(time, tslocal[ts], _fs, maxValue[ts], minValue[ts],
1, maxValueT[ts], meanValueT[ts], false);
trainingPair.OutputVector = new List<double>();
trainingPair.OutputVector = createOutputVector(time, tslocal[ts], _fs, maxValue[ts],
minValue[ts]);
//trainingPair.OutputVector.Add(_ts.timeseries[i] / maxValue[counter] - minValue[counter]);
if (trainingPair.InputVector != null)
trainingSet[ts].Pairs.Add(trainingPair);
}
}
//do training
if (trainingSet[ts].Pairs.Any())
{
var mlNN = new MultiLayersNN(trainingSet[ts].Pairs[0].InputVector.Count, trainingSet[ts].Pairs[0].OutputVector.Count, 4, new[] { trainingSet[ts].Pairs[0].InputVector.Count, 8, 8, trainingSet[ts].Pairs[0].OutputVector.Count });
mlNN.trainNetwork(trainingSet[ts], 1000, 0.1, true);
int errorsNumber = mlNN.absoluteErrors.Count;
for (int i = 0; i < errorsNumber; i++)
{
mlNN.absoluteErrors[i] = (mlNN.absoluteErrors[i] * (maxValue[ts] - minValue[ts]) + minValue[ts]);
}
if (mlnset.Count <= ts)
{
mlnset.Add((mlNN));
}
else
{
mlnset[ts] = mlNN;
}
}
}
}
/// <summary>
/// Тренировка модели.
/// </summary>
/// <param name="source">Подготовленный список значений для тренировки.
/// "Подготовленность" заключается в том, что каждое значение списка содержит все значения в соответствующее время суток.
/// Например, source[0] содержит все значения энергопотребления в 00:00, source[1] - в 00:15, и так далее. В данном примере длина списка будет равна 96.</param>
/// <param name="settings">Настройки прогнозирования.</param>
/// <param name="additionalTraining">Производится ли дообучение модели.</param>
public void Train(List<TimeSeries> source, ForecastSettings settings, bool additionalTraining)
{
var trainSet = new List<StructuredDataSet>();
var cvSet = new List<StructuredDataSet>();
for (int i = 0; i < timestampNumber; i++)
{
trainSet.Add(new StructuredDataSet());
cvSet.Add(new StructuredDataSet());
double valueMax = 0;
double valueMin = double.MaxValue;
for (int day = 0; day < source[i].timeseries.Count; day++)
{
var value = source[i].timeseries[day];
valueMax = Math.Max(value, valueMax);
valueMin = Math.Min(value, valueMin);
}
maxValues[i] = valueMax;
minValues[i] = valueMin;
trainSet[i].Pairs = new List<DataPair>();
//преобразуем то, что нужно в выборки данных.
//формируем trainingset
for (int time = settings.energyLags.Max(); time < source[i].timeseries.Count(); time++)
{
DataPair trainingPair = new DataPair();
trainingPair.InputVector = createInputVector(time, source[i],
settings, maxValues[i], minValues[i], 1);
trainingPair.OutputVector = createOutputVector(time, source[i],
maxValues[i], minValues[i]);
if(trainingPair.InputVector != null)
trainSet[i].Pairs.Add(trainingPair);
}
if (trainSet[i].Pairs.Any())
{
var neuralNetwork = new MultiLayersNN(
trainSet[i].Pairs[0].InputVector.Count,
trainSet[i].Pairs[0].OutputVector.Count,
4,
new[]
{
trainSet[i].Pairs[0].InputVector.Count,
8, 8,
trainSet[i].Pairs[0].OutputVector.Count
});
neuralNetwork.trainNetwork(trainSet[i], 1000, 0.1, true);
for (int e = 0; e < neuralNetwork.absoluteErrors.Count; e++)
{
//денормализация
neuralNetwork.absoluteErrors[e] = (neuralNetwork.absoluteErrors[e] * (maxValues[i] - minValues[i]) + minValues[i]);
}
if (networkSet.Count <= i)
networkSet.Add(neuralNetwork);
else
networkSet[i] = neuralNetwork;
}
}
}
protected override DataStreamUnit[] _Execute(DataStreamUnit[] input)
{
var value = input[0];
value.Values["Number"] = counter;
var waitUntil = DateTime.Parse(value.Values["EnergyLag"].ToString());
var dateTime = DateTime.Parse(value.Values["TimeStamp"].ToString());
if (dateTime > waitUntil)
{
List<double> absoluteErrors = new List<double>();
neuralNetwork = parameters["NeuralNetwork"] as MultiLayersNN;
if (neuralNetwork != null)
{
absoluteErrors = neuralNetwork.absoluteErrors;
}
var realValue = Double.Parse(value.Values["Value"].ToString());
var predictedValue = value.Values["PredictedValue"] != null ? Double.Parse(value.Values["PredictedValue"].ToString()) : (double?)null;
if (predictedValue != null)
{
var absoluteError = Math.Abs(predictedValue.Value - realValue);
if (realValue != 0)
{
counter++;
mapeSum += Math.Abs(absoluteError/realValue);
}
testCounter++;
value.Values["MAPE"] = mapeSum/testCounter;
}
D = (int)Math.Round(absoluteErrors.Count / 1.0, MidpointRounding.AwayFromZero);
value.Values["D"] = D;
double errorSum = 0;
for (int i = absoluteErrors.Count - D; i < absoluteErrors.Count; i++)
{
errorSum += absoluteErrors[i];
}
var mu = errorSum / D;
value.Values["mu"] = mu;
double sum = 0;
for (int i = makeNonNegative(absoluteErrors.Count - D) + 1; i < absoluteErrors.Count; i++)
{
sum += Math.Pow(absoluteErrors[i] - mu, 2);
}
var resultError = Math.Sqrt(sum) / D;
value.Values["Error"] = resultError;
}
else
{
value.Values["MAPE"] = null;
}
return input;
}