Page 1 of 12

Different Temperature Prediction Models for

Asphalt Concrete Pavement

Presenter: Saroj Pathak

PhD student,

Department of Civil and Environmental Engineering,

University of Hawaii at Manoa

spathak@hawaii.edu

Page 2 of 12

Introduction

➢ Mechanical properties of the pavement can vary

significantly with the magnitude of temperature

changes.

➢ Temperature variation of the asphalt pavement,

over a section, is mainly affected by the current and

past histories of the surrounding weather

conditions (time series-based problem) and

material characteristics of the pavement.

➢ In this project, different Artificial

Neural Network (ANN) based models

are developed to predict the

temperature of the asphalt pavement

section by considering climatic

conditions as the input.

➢ The models do not incorporate the

effects of material properties of the

pavement on the temperature variation.

Page 3 of 12

Objective

➢ To compare the performances and usefulness of the 4 different ANN based models

(Simple feedforward, 1D convolutional, SimpleRNN, and LSTM) to predict the

temperature of the asphalt concrete section at different depths.

➢ To outline possible future research to improve the model performances.

Page 4 of 12

Data Collection and Processing

➢ Temperature data collected for 8

different depth across a pavement

section, throughout the year (from Aug

2015-Aug 2016).

➢ Collected data merged with the weather

data (surface temperature, precipitation,

wind speed, cloud cover, humidity).

Page 5 of 12

Model Implementation

For feedforward (simple deep learning) model

• import keras

• from keras.models import Sequential

• from keras.layers import Dense

• model =Sequential( )

• model.add(Dense(20,input_dim=5,activation = 'relu'))

• model.add(Dense(50,activation = 'relu'))

• model.add(Dense(80,activation = 'relu'))

• model.add(Dense(50,activation = 'relu'))

• model.add(Dense(20,activation = 'relu'))

model.add(Dense(y.shape[1],activation = 'linear'))

• model.compile(loss='mean_squared_error',optimizer='Adam',metrics=['mea

n_squared_error'])

• model.summary()

For 1D Convolutional model

• model = Sequential()

• model.add(Conv1D(filters=64, kernel_size=10, activation='relu', input_shape

=(ntimesteps,num_features)))

• model.add(MaxPooling1D(pool_size=2))

• model.add(Conv1D(filters=64, kernel_size=10, activation='relu'))

• #model.add(MaxPooling1D(pool_size=2))

• model.add(Flatten())

• model.add(Dense(50, activation='relu'))

• model.add(Dense(8))

• model.compile(optimizer='adam', loss='mse')

• model.summary()

For SimpleRNN model

• import keras

• from keras.models import Sequential

• from tensorflow.keras.layers import SimpleRNN, Dropout, Dense

model = Sequential()

model.add(SimpleRNN(units=32, input_shape=(ntimesteps,num_fea

tures),stateful=False, activation='relu'))

model.add(Dense(8))

• model.compile(loss='mean_squared_error', optimizer='adam')

• model.summary()

For LSTM model

• import keras

• from keras.models import Sequential

• from tensorflow.keras.layers import LSTM, Dense, TimeDistributed

model = Sequential()

model.add(LSTM(units=16,input_shape=(ntimesteps,num_features),

return_sequences=True))

• model.add(LSTM(units=16,activation='relu'))

• model.add(Dense(8))

• model.compile(loss='mean_squared_error', optimizer='adam')

• model.summary()

Page 6 of 12

Model Performances

For Feedforward Model:

Page 7 of 12

Model Performances

For 1D Convolutional Model:

Page 8 of 12

Model Performances

For SimpleRNN Model:

Page 9 of 12

Model Performances

For LSTM Model:

Page 10 of 12

Result, Conclusion and Future Research

Result

➢ As expected, time series-based models outperformed the simple feedforward model (reason: the simple

feedforward model was only trained for 1 time step data).

➢ Within the time series-based models, Simple RNN and LSTM model were relatively more accurate compared

to 1D Convolution model although all of them were trained using same time steps of the data.

➢ LSTM model was slightly more accurate than simple RNN model.

Conclusion

➢ The LSTM or Simple RNN model can be used to predict pavement temperature profile with very good

accuracy.

Future Research

➢ Incorporating pavement thermal and material properties in the model training can enhance model

performances and usefulness.

Page 11 of 12

References

➢ Arangi, S.R.; Jain, R.K. Review Paper on Pavement Temperature Prediction Model for Indian Climatic Condition. Int. J. Innov.

Res. Adv. Eng. 2015, 2, 1–9.

➢ Matic, B.; Matic, D.; Cosic, D.; Sremac, S.; Tepic, G.; Ranitovic, P. A Model for the Pavement Temperature Prediction at Specified

Depth. Metalurgija 2013, 52, 505–508.

➢ Van Dam, T.J.; Harvey, J.T.; Muench, S.T.; Smith, K.D.; Snyder, M.B.; Al-Qadi, I.L.; Ozer, H.; Meijer, J.; Ram, P.V.; Roesler, J.R.; et al.

Towards Sustainable Pavement Systems; Applied Pavement Technology Inc.: Urbana, IL, USA, 2015.

Page 12 of 12

Questions ???

Thank You!