Main

LSTM Networks

Main.LSTMNetwork History

Hide minor edits - Show changes to output

January 15, 2020, at 02:44 PM by 147.46.252.163 -
Changed lines 7-8 from:
%width=300px%Attach:lstm_node.png
to:
(:html:)
<iframe width="560" height="315" src="https://www
.youtube.com/embed/LiBFV7ptm4M" frameborder="0" allow="accelerometer; autoplay; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe>
(:htmlend:)

Added lines 60-61:
%width=300px%Attach:lstm_node.png
Added lines 229-232:

(:html:)
<iframe width="560" height="315" src="https://www.youtube.com/embed/EC3QOvgRDSs" frameborder="0" allow="accelerometer; autoplay; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe>
(:htmlend:)
January 15, 2020, at 12:41 PM by 147.46.252.163 -
Changed lines 381-384 from:
plt.figure(figsize=(10,5))
plt.plot(train['Time'][window:],Yu,label='LSTM')
plt.plot(train['Time'][window:],Ym,label='Measured')
plt.plot(train['Q1'],label='heater')
to:
plt.figure(figsize=(10,6))
plt.subplot(2,1,1)
plt.plot(
train['Time'][window:],Yu,'r-',label='LSTM')
plt.plot(train['Time'][window:],Ym,'k--',label='Measured')
plt.ylabel('Temperature (°C)')
Changed lines 387-390 from:
plt.xlabel('Time (sec)'); plt.ylabel('Values')
to:
plt.subplot(2,1,2)
plt.plot(train['Q1'],label='heater (%)')
plt.legend()
plt.xlabel('Time (sec)'); plt.ylabel('Heater
')
Deleted line 392:
Changed line 420 from:
plt.figure(figsize=(10,5))
to:
plt.figure(figsize=(10,6))
Changed line 442 from:
plt.figure(figsize=(10,5))
to:
plt.figure(figsize=(10,6))
December 01, 2019, at 05:37 AM by 136.36.211.159 -
Changed line 386 from:
plt.xlabel('Time (sec)'); ylabel('Values')
to:
plt.xlabel('Time (sec)'); plt.ylabel('Values')
December 01, 2019, at 05:35 AM by 136.36.211.159 -
Changed lines 224-226 from:
Develop a model of the dynamic temperature response of the TCLab and compare the LSTM model prediction to a [[https://apmonitor.com/pdc/index.php/Main/TCLabSecondOrder|second-order linear differential equation solution]]. Use the [[Attach:tclab_dyn_data3.txt|4 hours of dynamic data from a TCLab]] (14400 data points = 1 second sample rate for 4 hours) or else generate new data (840 data points = 1 second sample rate for 14 min).

(:toggle hide tclab_data button show="Generate New TCLab Data
":)
to:
Develop a model of the dynamic temperature response of the TCLab and compare the LSTM model prediction to a [[https://apmonitor.com/pdc/index.php/Main/TCLabSecondOrder|second-order linear differential equation solution]]. Use the [[Attach:tclab_dyn_data3.txt|4 hours of dynamic data from a TCLab]] (14400 data points = 1 second sample rate for 4 hours) for training and generate new data (840 data points = 1 second sample rate for 14 min) for validation (see [[https://apmonitor.com/pdc/uploads/Main/tclab_data4.txt|sample validation data]]).

(:toggle hide tclab_data button show="Generate New TCLab Data for Validation
":)
Added lines 288-289:
The LSTM model is trained with the TCLab 4 hours data set for 10 epochs. The loss function decreases for the first few epochs and then does not significantly change after that.
Added lines 292-303:
The model predictions have good agreement with the measurements. The next steps are to perform validation to determine the predictive capability of the model on a different data set.

%width=550px%Attach:lstm_tclab_fit.png

The validation shows the performance on data that was not used in the training. In this case, a separate data file is generated versus splitting the training and test data.

%width=550px%Attach:lstm_tclab_validation.png

The forecast is generated by using the prior LSTM predictions to predict future temperatures. The measurements are only used for initializing the predictions and then predictions are used to predict following values.

%width=550px%Attach:lstm_tclab_forecast.png

Added line 310:
Changed lines 331-332 from:
# Load historic values of length 'window' and next real value for model
# Each time step uses last 'window' number of changes
to predict the next change
to:
# Each time step uses last 'window' to predict the next change
Changed lines 347-348 from:
model.add(LSTM(units=50, return_sequences=True, input_shape=(X.shape[1],X.shape[2])))
to:
model.add(LSTM(units=50, return_sequences=True, \
       
input_shape=(X.shape[1],X.shape[2])))
Changed lines 355-356 from:
model.compile(optimizer = 'adam', loss = 'mean_squared_error', metrics = ['accuracy'])
to:
model.compile(optimizer = 'adam', loss = 'mean_squared_error',\
           
metrics = ['accuracy'])
Added line 370:
plt.xlabel('epoch'); plt.ylabel('loss')
Added lines 385-386:
plt.legend()
plt.xlabel('Time (sec)'); ylabel('Values')
Changed lines 388-397 from:
plt.show()
(:sourceend:)

%width=550px%Attach:lstm_tclab_fit.png

'''TCLab LSTM Validation'''

%width=550px%Attach:lstm_tclab_validation.png

(:source lang=python:)
to:


# Load model
Changed lines 428-436 from:
(:sourceend:)

'''TCLab LSTM Forecast'''

The forecast is generated by using the prior LSTM predictions to predict future temperatures. The measurements are only used for initializing the predictions and then predictions are used to predict following values.

%width=550px%Attach:lstm_tclab_forecast.png

(:source lang=python:)
to:
Changed line 435 from:
      
to:
Added line 452:
December 01, 2019, at 05:21 AM by 136.36.211.159 -
Deleted line 293:
%matplotlib inline
Changed line 295 from:
to:
import time
December 01, 2019, at 05:14 AM by 136.36.211.159 -
Changed lines 224-232 from:
Develop a model of the dynamic temperature response of the TCLab and compare the LSTM model prediction to a [[https://apmonitor.com/pdc/index.php/Main/TCLabSecondOrder|second-order linear differential equation solution]]. Use the prior temperature values and heater values to predict the next temperature value. Show the model validation that predicts based on prior data as well as a forecast assessment where measurements are not used to generate the predictions.

(:toggle hide lstm button show="Show LSTM Solution":)
(:div id=lstm:)

'''Solution with LSTM Model'''

%width=500px%Attach:lstm_tclab.png
to:
Develop a model of the dynamic temperature response of the TCLab and compare the LSTM model prediction to a [[https://apmonitor.com/pdc/index.php/Main/TCLabSecondOrder|second-order linear differential equation solution]]. Use the [[Attach:tclab_dyn_data3.txt|4 hours of dynamic data from a TCLab]] (14400 data points = 1 second sample rate for 4 hours) or else generate new data (840 data points = 1 second sample rate for 14 min).

(:toggle hide tclab_data button show="Generate New TCLab Data":)
(:div id=tclab_data:)

%width=550px%Attach:tclab_data4.png
Changed lines 232-447 from:
to:
# generate new data
import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
import tclab
import time

n = 840  # Number of second time points (14 min)
tm = np.linspace(0,n,n+1) # Time values
lab = tclab.TCLab()
T1 = [lab.T1]
T2 = [lab.T2]
Q1 = np.zeros(n+1)
Q2 = np.zeros(n+1)
Q1[30:] = 35.0
Q1[270:] = 70.0
Q1[450:] = 10.0
Q1[630:] = 60.0
Q1[800:] = 0.0
for i in range(n):
    lab.Q1(Q1[i])
    lab.Q2(Q2[i])
    time.sleep(1)
    print(Q1[i],lab.T1)
    T1.append(lab.T1)
    T2.append(lab.T2)
lab.close()
# Save data file
data = np.vstack((tm,Q1,Q2,T1,T2)).T
np.savetxt('tclab_data.csv',data,delimiter=',',\
          header='Time,Q1,Q2,T1,T2',comments='')

# Create Figure
plt.figure(figsize=(10,7))
ax = plt.subplot(2,1,1)
ax.grid()
plt.plot(tm/60.0,T1,'r.',label=r'$T_1$')
plt.ylabel(r'Temp ($^oC$)')
ax = plt.subplot(2,1,2)
ax.grid()
plt.plot(tm/60.0,Q1,'b-',label=r'$Q_1$')
plt.ylabel(r'Heater (%)')
plt.xlabel('Time (min)')
plt.legend()
plt.savefig('tclab_data.png')
plt.show()
(:sourceend:)
(:divend:)

Use the measured temperature and heater values to predict the next temperature value with an LSTM model. Validate the model with a new data set in a predictive and forecast mode. The predictive mode predicts one step ahead while the forecast does not use temperature measurements to generate the predictions.

(:toggle hide lstm button show="Show LSTM Solution":)
(:div id=lstm:)

'''Solution with LSTM Model'''

%width=550px%Attach:lstm_tclab_loss.png

(:source lang=python:)
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
from sklearn.preprocessing import MinMaxScaler

# For LSTM model
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import LSTM
from keras.layers import Dropout
from keras.callbacks import EarlyStopping
from keras.models import load_model

# Load training data
file = 'https://apmonitor.com/do/uploads/Main/tclab_dyn_data3.txt'
train = pd.read_csv(file)

# Scale features
s1 = MinMaxScaler(feature_range=(-1,1))
Xs = s1.fit_transform(train[['T1','Q1']])

# Scale predicted value
s2 = MinMaxScaler(feature_range=(-1,1))
Ys = s2.fit_transform(train[['T1']])

# Load historic values of length 'window' and next real value for model
# Each time step uses last 'window' number of changes to predict the next change
window = 70
X = []
Y = []
for i in range(window,len(Xs)):
    X.append(Xs[i-window:i,:])
    Y.append(Ys[i])

# Reshape data to format accepted by LSTM
X, Y = np.array(X), np.array(Y)

# create and train LSTM model

# Initialize LSTM model
model = Sequential()

model.add(LSTM(units=50, return_sequences=True, input_shape=(X.shape[1],X.shape[2])))
model.add(Dropout(0.2))
model.add(LSTM(units=50, return_sequences=True))
model.add(Dropout(0.2))
model.add(LSTM(units=50))
model.add(Dropout(0.2))
model.add(Dense(units=1))
model.compile(optimizer = 'adam', loss = 'mean_squared_error', metrics = ['accuracy'])

# Allow for early exit
es = EarlyStopping(monitor='loss',mode='min',verbose=1,patience=10)

# Fit (and time) LSTM model
t0 = time.time()
history = model.fit(X, Y, epochs = 10, batch_size = 250, callbacks=[es], verbose=1)
t1 = time.time()
print('Runtime: %.2f s' %(t1-t0))

# Plot loss
plt.figure(figsize=(8,4))
plt.semilogy(history.history['loss'])
plt.savefig('tclab_loss.png')
model.save('model.h5')

# Verify the fit of the model
Yp = model.predict(X)

# un-scale outputs
Yu = s2.inverse_transform(Yp)
Ym = s2.inverse_transform(Y)

plt.figure(figsize=(10,5))
plt.plot(train['Time'][window:],Yu,label='LSTM')
plt.plot(train['Time'][window:],Ym,label='Measured')
plt.plot(train['Q1'],label='heater')
plt.savefig('tclab_fit.png')
plt.show()
(:sourceend:)

%width=550px%Attach:lstm_tclab_fit.png

'''TCLab LSTM Validation'''

%width=550px%Attach:lstm_tclab_validation.png

(:source lang=python:)
v = load_model('model.h5')
# Load training data
test = pd.read_csv('https://apmonitor.com/pdc/uploads/Main/tclab_data4.txt')

Xt = test[['T1','Q1']].values
Yt = test[['T1']].values

Xts = s1.transform(Xt)
Yts = s2.transform(Yt)

Xti = []
Yti = []
for i in range(window,len(Xts)):
    Xti.append(Xts[i-window:i,:])
    Yti.append(Yts[i])

# Reshape data to format accepted by LSTM
Xti, Yti = np.array(Xti), np.array(Yti)

# Verify the fit of the model
Ytp = model.predict(Xti)

# un-scale outputs
Ytu = s2.inverse_transform(Ytp)
Ytm = s2.inverse_transform(Yti)

plt.figure(figsize=(10,5))
plt.subplot(2,1,1)
plt.plot(test['Time'][window:],Ytu,'r-',label='LSTM Predicted')
plt.plot(test['Time'][window:],Ytm,'k--',label='Measured')
plt.legend()
plt.ylabel('Temperature (°C)')
plt.subplot(2,1,2)
plt.plot(test['Time'],test['Q1'],'b-',label='Heater')
plt.xlabel('Time (sec)'); plt.ylabel('Heater (%)')
plt.legend()
plt.savefig('tclab_validate.png')
(:sourceend:)

'''TCLab LSTM Forecast'''

The forecast is generated by using the prior LSTM predictions to predict future temperatures. The measurements are only used for initializing the predictions and then predictions are used to predict following values.

%width=550px%Attach:lstm_tclab_forecast.png

(:source lang=python:)
# Using predicted values to predict next step
Xtsq = Xts.copy()
for i in range(window,len(Xtsq)):
    Xin = Xtsq[i-window:i].reshape((1, window, 2))
    Xtsq[i][0] = v.predict(Xin)
    Yti[i-window] = Xtsq[i][0]
       
#Ytu = (Yti - s2.min_[0])/s2.scale_[0]
Ytu = s2.inverse_transform(Yti)

plt.figure(figsize=(10,5))
plt.subplot(2,1,1)
plt.plot(test['Time'][window:],Ytu,'r-',label='LSTM Predicted')
plt.plot(test['Time'][window:],Ytm,'k--',label='Measured')
plt.legend()
plt.ylabel('Temperature (°C)')
plt.subplot(2,1,2)
plt.plot(test['Time'],test['Q1'],'b-',label='Heater')
plt.xlabel('Time (sec)'); plt.ylabel('Heater (%)')
plt.legend()
plt.savefig('tclab_forecast.png')
plt.show()
November 30, 2019, at 05:02 AM by 136.36.211.159 -
Added lines 229-230:
'''Solution with LSTM Model'''
Added lines 240-241:

'''Solution with 2nd-Order Model'''
November 30, 2019, at 05:00 AM by 136.36.211.159 -
Changed lines 241-242 from:
{$\min \sum_{i=1}^{n} \left(T_{C1,meas}-T_{C1,pred}\right)^2$}
to:
{$\min \sum_{i=1}^{n} \left(T_{C1,meas,i}-T_{C1,pred,i}\right)^2$}
Changed line 245 from:
{$\tau_2 \frac{dT_{C1}}{dt} + \left(T_{C1}\right) = T_{H1}$}
to:
{$\tau_2 \frac{dT_{C1}}{dt} + T_{C1} = T_{H1}$}
November 30, 2019, at 04:59 AM by 136.36.211.159 -
Changed line 241 from:
{$\min \sum_{i=1}^{n} \left(T_{C1,meas}-T_{C1,pred}\right)^2
to:
{$\min \sum_{i=1}^{n} \left(T_{C1,meas}-T_{C1,pred}\right)^2$}
November 30, 2019, at 04:59 AM by 136.36.211.159 -
Changed lines 239-241 from:
A second order model is aligned with data by adjusting unknown parameters {`K_1`}, {`\tau_1`}, and {`\tau_2`} to minimize the sum of squared errors.
to:
A second order model is aligned with data by adjusting unknown parameters to minimize the sum of squared errors. The adjusted parameters are {`K_1`}, {`\tau_1`}, and {`\tau_2`}.

{$\min \sum_{i=1}^{n} \left(T_{C1,meas}-T_{C1,pred}\right)^2
November 30, 2019, at 04:56 AM by 136.36.211.159 -
Changed line 226 from:
(:toggle hide lstm button show="Show LSTM Keras TCLab Code":)
to:
(:toggle hide lstm button show="Show LSTM Solution":)
Added lines 233-374:
(:sourceend:)
(:divend:)

(:toggle hide second_order button show="Show 2nd-Order Solution":)
(:div id=second_order:)

A second order model is aligned with data by adjusting unknown parameters {`K_1`}, {`\tau_1`}, and {`\tau_2`} to minimize the sum of squared errors.

{$\tau_1 \frac{dT_{H1}}{dt} + \left(T_{H1}-T_a\right) = K_1 \, Q_1$}

{$\tau_2 \frac{dT_{C1}}{dt} + \left(T_{C1}\right) = T_{H1}$}

The first 3000 data points train the model and the next 6000 data points validate the model predictions on data that has not been used for training.

'''Training 2nd-Order Model'''

%width=500px%Attach:tclab_2nd_order_train.png

'''Validating 2nd-Order Model'''

%width=500px%Attach:tclab_2nd_order_validate.png

(:source lang=python:)
import numpy as np
import matplotlib.pyplot as plt
from gekko import GEKKO
import pandas as pd

file = 'https://apmonitor.com/do/uploads/Main/tclab_dyn_data3.txt'
data = pd.read_csv(file)

# subset for training
n = 3000
tm = data['Time'][0:n].values
Q1s = data['Q1'][0:n].values
T1s = data['T1'][0:n].values

m = GEKKO()
m.time = tm

# Parameters to Estimate
K1 = m.FV(value=0.5,lb=0.1,ub=1.0)
tau1 = m.FV(value=150,lb=50,ub=250)
tau2 = m.FV(value=15,lb=10,ub=20)
K1.STATUS = 1
tau1.STATUS = 1
tau2.STATUS = 1

# Model Inputs
Q1 = m.Param(value=Q1s)
Ta = m.Param(value=23.0) # degC
T1m = m.Param(T1s)

# Model Variables
TH1 = m.Var(value=T1s[0])
TC1 = m.Var(value=T1s)

# Objective Function
m.Minimize((T1m-TC1)**2)

# Equations
m.Equation(tau1 * TH1.dt() + (TH1-Ta) == K1*Q1)
m.Equation(tau2 * TC1.dt()  + TC1 == TH1)

# Global Options
m.options.IMODE  = 5 # MHE
m.options.EV_TYPE = 2 # Objective type
m.options.NODES  = 2 # Collocation nodes
m.options.SOLVER  = 3 # IPOPT

# Predict Parameters and Temperatures
m.solve()

# Create plot
plt.figure(figsize=(10,7))

ax=plt.subplot(2,1,1)
ax.grid()
plt.plot(tm,T1s,'ro',label=r'$T_1$ measured')
plt.plot(tm,TC1.value,'k-',label=r'$T_1$ predicted')
plt.ylabel('Temperature (degC)')
plt.legend(loc=2)
ax=plt.subplot(2,1,2)
ax.grid()
plt.plot(tm,Q1s,'b-',label=r'$Q_1$')
plt.ylabel('Heater (%)')
plt.xlabel('Time (sec)')
plt.legend(loc='best')

# Print optimal values
print('K1: ' + str(K1.newval))
print('tau1: ' + str(tau1.newval))
print('tau2: ' + str(tau2.newval))

# Save and show figure
plt.savefig('tclab_2nd_order_fit.png')


# Validation
tm = data['Time'][n:3*n].values
Q1s = data['Q1'][n:3*n].values
T1s = data['T1'][n:3*n].values

v = GEKKO()
v.time = tm

# Parameters to Estimate
K1 = K1.newval
tau1 = tau1.newval
tau2 = tau2.newval
Q1 = v.Param(value=Q1s)
Ta = v.Param(value=23.0) # degC
TH1 = v.Var(value=T1s[0])
TC1 = v.Var(value=T1s[0])
v.Equation(tau1 * TH1.dt() + (TH1-Ta) == K1*Q1)
v.Equation(tau2 * TC1.dt()  + TC1 == TH1)
v.options.IMODE  = 4 # Simulate
v.options.NODES  = 2 # Collocation nodes
v.options.SOLVER  = 1

# Predict Parameters and Temperatures
v.solve(disp=True)

# Create plot
plt.figure(figsize=(10,7))

ax=plt.subplot(2,1,1)
ax.grid()
plt.plot(tm,T1s,'ro',label=r'$T_1$ measured')
plt.plot(tm,TC1.value,'k-',label=r'$T_1$ predicted')
plt.ylabel('Temperature (degC)')
plt.legend(loc=2)
ax=plt.subplot(2,1,2)
ax.grid()
plt.plot(tm,Q1s,'b-',label=r'$Q_1$')
plt.ylabel('Heater (%)')
plt.xlabel('Time (sec)')
plt.legend(loc='best')

# Save and show figure
plt.savefig('tclab_2nd_order_validate.png')
plt.show()
November 30, 2019, at 03:32 AM by 136.36.211.159 -
Changed line 137 from:
(:toggle hide keras button show="Show All Keras Example Code":)
to:
(:toggle hide keras button show="Show Sine Keras Example Code":)
Changed lines 226-227 from:
(:toggle hide gekko button show="Show Python Source":)
(:div id=gekko:)
to:
(:toggle hide lstm button show="Show LSTM Keras TCLab Code":)
(:div id=lstm:)
November 29, 2019, at 04:56 AM by 136.36.211.159 -
Changed line 224 from:
Develop a model of the dynamic temperature response of the TCLab and compare the LSTM model prediction to a [[https://apmonitor.com/pdc/index.php/Main/TCLabSecondOrder|second-order linear differential equation solution]]. Use the prior temperature values and heater values to predict the next temperature value. Show the model validation that predicts based on prior data as well as a predictive assessment where measurements are not used to generate the predictions.
to:
Develop a model of the dynamic temperature response of the TCLab and compare the LSTM model prediction to a [[https://apmonitor.com/pdc/index.php/Main/TCLabSecondOrder|second-order linear differential equation solution]]. Use the prior temperature values and heater values to predict the next temperature value. Show the model validation that predicts based on prior data as well as a forecast assessment where measurements are not used to generate the predictions.
November 29, 2019, at 04:38 AM by 136.36.211.159 -
Changed line 11 from:
!!!! Data Preparation
to:
'''Data Preparation'''
November 29, 2019, at 04:32 AM by 136.36.211.159 -
Changed lines 89-90 from:
'''LSTM Model Validation'''
to:
'''LSTM Prediction Validation'''
Changed lines 117-119 from:
When performing the validation it is also important to determine how the model performs with without measurements when it uses prior predictions (not measurements) to predict the next outcome. This is important to determine how well the model performs in a predictive application such as model predictive control where the model is projected forward minutes to hours to determine the sequence of optimal manipulated variable moves and possible future constraint violation.
to:
'''LSTM Forecast Validation'''

When performing the validation it is also important to determine how the model performs with without measurements when it uses prior predictions
to predict the next outcome. This is important to determine how well the model performs in a predictive application such as model predictive control where the model is projected forward over the control horizon to determine the sequence of optimal manipulated variable moves and possible future constraint violation. Generating predictions without measurement feedback is a forecast.
November 29, 2019, at 04:20 AM by 136.36.211.159 -
Changed line 222 from:
Repeat the above exercise with the TCLab and compare the LSTM model prediction to a second order linear differential equation model. Use the prior temperature values and heater values to predict the next temperature value. Show the model validation with one-step forward as well as a predictive assessment where measurements are not used to generate the predictions.
to:
Develop a model of the dynamic temperature response of the TCLab and compare the LSTM model prediction to a [[https://apmonitor.com/pdc/index.php/Main/TCLabSecondOrder|second-order linear differential equation solution]]. Use the prior temperature values and heater values to predict the next temperature value. Show the model validation that predicts based on prior data as well as a predictive assessment where measurements are not used to generate the predictions.
November 29, 2019, at 04:12 AM by 136.36.211.159 -
Changed line 7 from:
%width=350px%Attach:lstm_node.png
to:
%width=300px%Attach:lstm_node.png
November 29, 2019, at 04:08 AM by 136.36.211.159 -
Changed line 9 from:
The LSTM is trained (parameters adjusted) with an input window of prior data and minimized difference between the predicted and next measured value. Sequential methods predict just one next value based on the window of prior data. When there is contextual data (before and after) the desired prediction point, a [[https://en.wikipedia.org/wiki/Convolutional_neural_network|Convolutional Neural Networks (CNN)]] may [[https://towardsdatascience.com/the-fall-of-rnn-lstm-2d1594c74ce0|improve performance with fewer resources to train and deploy the network]].
to:
The LSTM is trained (parameters adjusted) with an input window of prior data and minimized difference between the predicted and next measured value. Sequential methods predict just one next value based on the window of prior data. When there is contextual data (before and after) the desired prediction point, a [[https://en.wikipedia.org/wiki/Convolutional_neural_network|Convolutional Neural Network (CNN)]] may [[https://towardsdatascience.com/the-fall-of-rnn-lstm-2d1594c74ce0|improve performance with fewer resources to train and deploy the network]].
November 28, 2019, at 07:45 PM by 136.36.211.159 -
Changed line 224 from:
(:toggle hide gekko button show="Show Python Gekko Source":)
to:
(:toggle hide gekko button show="Show Python Source":)
November 28, 2019, at 07:44 PM by 136.36.211.159 -
Changed line 9 from:
The LSTM is trained (parameters adjusted) with an input window of prior data and minimized difference between the predicted and next measured value. Sequential methods predict just one next value based on the window of prior data.
to:
The LSTM is trained (parameters adjusted) with an input window of prior data and minimized difference between the predicted and next measured value. Sequential methods predict just one next value based on the window of prior data. When there is contextual data (before and after) the desired prediction point, a [[https://en.wikipedia.org/wiki/Convolutional_neural_network|Convolutional Neural Networks (CNN)]] may [[https://towardsdatascience.com/the-fall-of-rnn-lstm-2d1594c74ce0|improve performance with fewer resources to train and deploy the network]].
November 28, 2019, at 07:36 PM by 136.36.211.159 -
Deleted lines 19-27:

Attach:lstm_tutorial.mp4

(:html:)
<video controls autoplay loop>
  <source src="/do/uploads/Main/lstm_tutorial.mp4" type="video/mp4">
  Your browser does not support the video tag.
</video>
(:htmlend:)
November 28, 2019, at 07:32 PM by 136.36.211.159 -
Deleted lines 68-69:
%width=550px%Attach:lstm_sine_loss.png
Changed line 87 from:
%width=550px%Attach:lstm_sine_fit.png
to:
%width=550px%Attach:lstm_sine_loss.png
November 28, 2019, at 07:30 PM by 136.36.211.159 -
Changed lines 69-70 from:
%width=550px%lstm_sine_loss.png
to:
%width=550px%Attach:lstm_sine_loss.png
Changed lines 89-90 from:
%width=550px%lstm_sine_fit.png
to:
%width=550px%Attach:lstm_sine_fit.png
Changed lines 104-105 from:
%width=550px%lstm_sine_pred1.png
to:
%width=550px%Attach:lstm_sine_pred1.png
Changed line 130 from:
%width=550px%lstm_sine_pred2.png
to:
%width=550px%Attach:lstm_sine_pred2.png
November 28, 2019, at 07:29 PM by 136.36.211.159 -
Changed line 21 from:
lstm_tutorial.mp4
to:
Attach:lstm_tutorial.mp4
November 28, 2019, at 07:28 PM by 136.36.211.159 -
Added lines 1-243:
(:title LSTM Networks:)
(:keywords sequential processing, artificial intelligence, machine learning, tutorial, tensorflow, keras, gekko:)
(:description Long-Short Term Memory (LSTM), Recurrent Neural Networks, and other sequential processing methods consider a window of data to make a future prediction.:)

LSTM (Long Short Term Memory) networks are a special type of RNN (Recurrent Neural Network) that is structured to remember and predict based on long-term dependencies that are trained with time-series data. An LSTM repeating module has four interacting components.

%width=350px%Attach:lstm_node.png

The LSTM is trained (parameters adjusted) with an input window of prior data and minimized difference between the predicted and next measured value. Sequential methods predict just one next value based on the window of prior data.

!!!! Data Preparation

Data preparation for LSTM networks involves consolidation, cleansing, separating the input window and output, scaling, and data division for training and validation.

* Consolidation - consolidation is the process of combining disparate data (Excel spreadsheet, PDF report, database, cloud storage) into a single repository.
* Data Cleansing - bad data should be removed and may include outliers, missing entries, failed sensors, or other types of missing or corrupted information.
* Inputs and Outputs - data is separated into inputs (prior time-series window) and outputs (predicted next value). The inputs are fed into a series of functions to produce the output prediction. The squared difference between the predicted output and the measured output is a typical loss (objective) function for fitting. 
* Scaling - scaling all data (inputs and outputs) to a range of 0-1 can improve the training process.
* Training and Validation - data is divided into training (e.g. 80%) and validation (e.g. 20%) sets so that the model fit can be evaluated independently of the training. Cross-validation is an approach to divide the training data into multiple sets that are fit separately. The parameter consistency is compared between the multiple models.

lstm_tutorial.mp4

(:html:)
<video controls autoplay loop>
  <source src="/do/uploads/Main/lstm_tutorial.mp4" type="video/mp4">
  Your browser does not support the video tag.
</video>
(:htmlend:)

'''Data Generation and Preparation'''

(:source lang=python:)
import numpy as np
import matplotlib.pyplot as plt

# Generate data
n = 500
t = np.linspace(0,20.0*np.pi,n)
X = np.sin(t) # X is already between -1 and 1, scaling normally needed
(:sourceend:)

Once the data is created, it is converted to a form that can be used by Keras and Tensorflow for training and prediction.

(:source lang=python:)
# Set window of past points for LSTM model
window = 10

# Split 80/20 into train/test data
last = int(n/5.0)
Xtrain = X[:-last]
Xtest = X[-last-window:]

# Store window number of points as a sequence
xin = []
next_X = []
for i in range(window,len(Xtrain)):
    xin.append(Xtrain[i-window:i])
    next_X.append(Xtrain[i])
   
# Reshape data to format for LSTM
xin, next_X = np.array(xin), np.array(next_X)
xin = xin.reshape(xin.shape[0], xin.shape[1], 1)
(:sourceend:)

'''LSTM Model Build'''

An LSTM network relates the input data window to outputs with layers. Instead of just one layer, LSTMs often have multiple layers.

%width=550px%lstm_sine_loss.png

(:source lang=python:)
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import LSTM
from keras.layers import Dropout

# Initialize LSTM model
m = Sequential()
m.add(LSTM(units=50, return_sequences=True, input_shape=(xin.shape[1],1)))
m.add(Dropout(0.2))
m.add(LSTM(units=50))
m.add(Dropout(0.2))
m.add(Dense(units=1))
m.compile(optimizer = 'adam', loss = 'mean_squared_error')
(:sourceend:)

'''LSTM Model Training'''

%width=550px%lstm_sine_fit.png

(:source lang=python:)
# Fit LSTM model
history = m.fit(xin, next_X, epochs = 50, batch_size = 50,verbose=0)

plt.figure()
plt.ylabel('loss'); plt.xlabel('epoch')
plt.semilogy(history.history['loss'])
(:sourceend:)

'''LSTM Model Validation'''

The validation test set assesses the ability of the neural network to predict based on new conditions that were not part of the training set. The validation is performed with the last 20% of the data that was separated from the beginning 80% of data.

%width=550px%lstm_sine_pred1.png

(:source lang=python:)
# Store "window" points as a sequence
xin = []
next_X1 = []
for i in range(window,len(Xtest)):
    xin.append(Xtest[i-window:i])
    next_X1.append(Xtest[i])
   
# Reshape data to format for LSTM
xin, next_X1 = np.array(xin), np.array(next_X1)
xin = xin.reshape((xin.shape[0], xin.shape[1], 1))

# Predict the next value (1 step ahead)
X_pred = m.predict(xin)

# Plot prediction vs actual for test data
plt.figure()
plt.plot(X_pred,':',label='LSTM')
plt.plot(next_X1,'--',label='Actual')
plt.legend()
(:sourceend:)

When performing the validation it is also important to determine how the model performs with without measurements when it uses prior predictions (not measurements) to predict the next outcome. This is important to determine how well the model performs in a predictive application such as model predictive control where the model is projected forward minutes to hours to determine the sequence of optimal manipulated variable moves and possible future constraint violation.

%width=550px%lstm_sine_pred2.png

(:source lang=python:)
# Using predicted values to predict next step
X_pred = Xtest.copy()
for i in range(window,len(X_pred)):
    xin = X_pred[i-window:i].reshape((1, window, 1))
    X_pred[i] = m.predict(xin)

# Plot prediction vs actual for test data
plt.figure()
plt.plot(X_pred[window:],':',label='LSTM')
plt.plot(next_X1,'--',label='Actual')
plt.legend()
(:sourceend:)

(:toggle hide keras button show="Show All Keras Example Code":)
(:div id=keras:)
(:source lang=python:)
import numpy as np
import matplotlib.pyplot as plt
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import LSTM
from keras.layers import Dropout

# Generate data
n = 500
t = np.linspace(0,20.0*np.pi,n)
X = np.sin(t) # X is already between -1 and 1, scaling normally needed

# Set window of past points for LSTM model
window = 10

# Split 80/20 into train/test data
last = int(n/5.0)
Xtrain = X[:-last]
Xtest = X[-last-window:]

# Store window number of points as a sequence
xin = []
next_X = []
for i in range(window,len(Xtrain)):
    xin.append(Xtrain[i-window:i])
    next_X.append(Xtrain[i])
   
# Reshape data to format for LSTM
xin, next_X = np.array(xin), np.array(next_X)
xin = xin.reshape(xin.shape[0], xin.shape[1], 1)

# Initialize LSTM model
m = Sequential()
m.add(LSTM(units=50, return_sequences=True, input_shape=(xin.shape[1],1)))
m.add(Dropout(0.2))
m.add(LSTM(units=50))
m.add(Dropout(0.2))
m.add(Dense(units=1))
m.compile(optimizer = 'adam', loss = 'mean_squared_error')

# Fit LSTM model
history = m.fit(xin, next_X, epochs = 50, batch_size = 50,verbose=0)

plt.figure()
plt.ylabel('loss'); plt.xlabel('epoch')
plt.semilogy(history.history['loss'])

# Store "window" points as a sequence
xin = []
next_X1 = []
for i in range(window,len(Xtest)):
    xin.append(Xtest[i-window:i])
    next_X1.append(Xtest[i])
   
# Reshape data to format for LSTM
xin, next_X1 = np.array(xin), np.array(next_X1)
xin = xin.reshape((xin.shape[0], xin.shape[1], 1))

# Predict the next value (1 step ahead)
X_pred = m.predict(xin)

# Plot prediction vs actual for test data
plt.figure()
plt.plot(X_pred,':',label='LSTM')
plt.plot(next_X1,'--',label='Actual')
plt.legend()

# Using predicted values to predict next step
X_pred = Xtest.copy()
for i in range(window,len(X_pred)):
    xin = X_pred[i-window:i].reshape((1, window, 1))
    X_pred[i] = m.predict(xin)

# Plot prediction vs actual for test data
plt.figure()
plt.plot(X_pred[window:],':',label='LSTM')
plt.plot(next_X1,'--',label='Actual')
plt.legend()
plt.show()
(:sourceend:)
(:divend:)

!!!! Exercise

Repeat the above exercise with the TCLab and compare the LSTM model prediction to a second order linear differential equation model. Use the prior temperature values and heater values to predict the next temperature value. Show the model validation with one-step forward as well as a predictive assessment where measurements are not used to generate the predictions.

(:toggle hide gekko button show="Show Python Gekko Source":)
(:div id=gekko:)

%width=500px%Attach:lstm_tclab.png

(:source lang=python:)

(:sourceend:)
(:divend:)