Skip to main content
Have a personal or library account? Click to login
Prediction of temporal atmospheric boundary layer height using long short-term memory network Cover

Prediction of temporal atmospheric boundary layer height using long short-term memory network

Tellus
Volume 73 (2021): Issue 1
By: ,   and    
Open Access
|Jan 2021
Figures & tables

Figures & Tables

Fig. 1.

Monsoon Map of India and SODAR system installed at Delhi (Courtesy: www.mapindia.com).

Fig. 2.

24 hours echogram of SODAR system.

Table 1.

Designing specifications of CSIR-NPL monostatic SODAR.

Transmitted power (electrcical)90 WattsTransmitted power (acoustical)15 WattsPulse width100 msPulse repetition period4 secOperational range1000 mReceiver bandwidth50 HzFrequency of operation2250 HzAcoustic velocity340 m/s (average)Receiver Gain80 dBTransmit-receive antennaParabolic reflector dish surrounded by conical acoustic cuffReceiver area2.5 sq. mPreamplifier sensitivityThe fraction of a micro-Volt
Fig. 3.

Network architecture.

Fig. 4.

Block diagram of the LSTM ABL height model.

Table 2.

LSTM architecture parameters.

Network parametersValueGradient decay factor0.9Squared gradient decay factor0.9990Initial learn rate0.005Learn rate schedulePiecewiseLearn rate drop factor0.20Learn rate drop period125aGradient threshold methodI2 normGradient threshold1Verbose frequency50Validation frequency50ShuffleonceSequence lengthLongestState activation functiontanhGate activation functionsigmoidInput weight initializerglorotRecurrent weight learn initializerorthogonalBias initializerUnit-forget-gateBias learns rate factor1

a Drop the learning rate after 125 epochs by a factor of 0.2.

Fig. 5.

Training progress with hidden layer 32 and Epoch 500.

Fig. 6.

Prediction result LSTM Network update with observed values (Prediction-2), hidden layer 32 and maximum epochs 500.

Table 3.

Comparison of accuracy of LSTM model.

Hidden layerPrediction-1Prediction-2RMSE1rRMSE1MAE1MAPE1RMSE2rRMSE2MAE2MAPE2Max epochs − 5002790.324.69617.572.37315.699.86197.4117.115317.839.93199.6120.22212.486.64133.9912.7410270.588.4518120.1214.716.7130.2412.2920250.937.84175.3618.27202.136.31121.8611.4325289.349.04197.8520.47204.766.39120.5911.4328290.059.06198.8922.33205.316.41118.3111.2530235.47.35166.9117.31192.016114.0110.9632220.317.33167.1217.3193.815.95113.7210.6235272.518.51194.4719.56197.156.16117.5210.8750264.618.26186.8419.27192.486.01112.2110.54100291.149.09205.5221.92198.866.21116.0611.11Max epochs − 25032373.8911.68250.2024.37193.666.05115.9910.90Max epochs − 75032271.668.48189.6220.68193.356.04112.0810.77

a Red colour row shows best prediction result.

Fig. 7.

Comparison of Line Plot of Observed ABL Height and Predication ABL Height.

Table 4.

Statistical analysis of LSTM model with hidden layer 32 and epoch 500.

ParameterTraining dataTest dataPrediction-1 dataPrediction-2 dataMean737837849852Median430551547400Kurtosis–1.22–1.67–1.66–1.71Skewness0.560.350.320.26Coefficient of variation0.800.680.720.78
Table 5.

Comparison between NAR and LSTM Model.

NARPrediction-1Prediction-2RMSE259.65220.31193.85MAE164.97167.12113.72
Table 6.

Monthly ABL height variation during different hours.

MonthMaximum ABL height (m)Corresponding hour of maximum ABL height (hrs)Average ABL height (m) duringDay time (09:00–18:00 hr)Remaining hours (19:00–08:00 hr)Diurnal average (m)December125513:00795170435January112013:00710200415February156513:001125265620March136012:001035185545April143513:001145225615May174512:001480350810June163513:001475325795July152012:001380325760August141012:001305260690September148514:001280190645October144512:00995225550November140511:00850230490
Fig. 8.

Annual ABL height Temporal and Monthly variation.

Fig. 9.

Prediction result LSTM Network update with predicted values (Prediction-1), hidden layer 32 and maximum epochs 500.

Fig. 10.

Box plot of temporal seasonal ABL height.

Fig. 11.

Seasonal ABL height Prediction result from the update network state with predicted values (Prediction-1).

Table 7.

Comparison of seasonal prediction of ABL height.

Data pointHidden layers – 32; Max Epochs – 500NAR modelsPrediction-1Prediction-2RMSE1rRMSE1MAE1MAPE1RMSE2rRMSE2MAE2MAPE2RMSEAnnual6984329.5510.29200.1323.35187.715.86118.7817.04261.80Winter1776247.257.72181.8319.59187.775.86125.8113.19307.33Pre-Monsoon2016266.308.32182.9920.02181.185.66116.2411.04245.91Monsoon2088300.479.38187.1415.62255.457.98151.7112.74281.25Post-Monsoon1104239.707.49163.7325.15178.925.59118.5715.17250.94
DOI: https://doi.org/10.1080/16000870.2021.1926132 | Journal eISSN: 3035-9554
Journal RSS Feed
Language: English
Page range: 1926132 - 1926132
Published on: Jan 1, 2021
Published by: Stockholm University Press
In partnership with: Paradigm Publishing Services
Keywords:
atmospheric boundary layer height,
LSTM network,
SODAR,
deep learning

© 2021 Nishant Kumar, Kirti Soni, Ravinder Agarwal, published by Stockholm University Press
This work is licensed under the Creative Commons Attribution 4.0 License.

Volume 73 (2021): Issue 1
Previous article
Next article