Hybrid Deep Learning for Week-Ahead Evapotranspiration Forecasting

doi:10.21203/rs.3.rs-424493/v1

Download PDF

Research Article

Hybrid Deep Learning for Week-Ahead Evapotranspiration Forecasting

https://doi.org/10.21203/rs.3.rs-424493/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Reference evapotranspiration (ET) is an integral hydrological factor in soil-plant-atmospheric water balance studies and the management of drought events. This paper proposes a new hybrid-deep learning (DL) approach, combining Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU) along with Ant Colony Optimization (ACO), for a multi-step (week 1 to 4) daily-ET forecast. The method also assimilates a comprehensive dataset with 52 diverse predictors, i.e., satellite-derived Moderate Resolution Imaging Spectroradiometer (MODIS), ground-based datasets from Scientific Information for Landowners (SILO) and synoptic-scale climate indices (CI). To develop a vigorous CNN-GRU model, a feature selection stage entails the Ant Colony Optimization (ACO) method implemented to improve the ET forecast model for the three selected sites in Australian Murray Darling Basin. The results demonstrate excellent forecasting capability of the hybrid CNN-GRU model against the counterpart benchmark models, evidenced by a relatively small mean absolute error and high efficiency. Overall, this study shows that the proposed hybrid CNN-GRU model successfully apprehends the complex and non-linear relationships between predictor variables and the daily ET.

Environmental Chemistry

Environmental Engineering

Convolutional neural network

gated recurrent unit

hybrid-deep learning

ET forecasting.

Table 1: Experimental Datasets Description. Note: The ID is the Station’s ID of BOM, Australia owned stations.

Datasets	SILO Database	MODIS-Satellite	Climate Mode Indices
Data type	Multivariable time series
Intervals	Daily
Stations	Menindee (ID: 047019): 32.39^οS, 142.42^οE Fairfield (ID: 066137): 35.53^οS, 144.97^οE Gabo Island (ID: 084016): 37.57^οS, 149.92^οE
Time Span	01/02/2003 to 31/03/2020
Number	9	35	13

Table 2 The RRMSE (%) error of CNN-GRU and other baseline models using single data sets

*Models*	*MODIS Satellite*				*SILO Data*				*Climate Mode Indices*
*Models*	*Week 1*	*Week 2*	*Week 3*	*Week 4*	*Week 1*	*Week 2*	*Week 3*	*Week 4*	*Week 1*	*Week 2*	*Week 3*	*Week 4*
CNN-GRU	1.075	1.375	1.675	1.975	1.575	1.875	2.175	2.475	1.945	2.245	2.545	2.845
CNN-LSTM	1.090	1.390	1.690	1.990	1.590	1.890	2.190	2.490	1.960	2.260	2.560	2.860
GRU	1.179	1.479	1.779	2.079	1.679	1.979	2.279	2.579	2.049	2.349	2.649	2.949
LSTM	1.425	1.725	2.025	2.325	1.925	2.225	2.525	2.825	2.295	2.595	2.895	3.195
RNN	1.518	1.818	2.118	2.418	2.018	2.318	2.618	2.918	2.388	2.688	2.988	3.288
DTree	1.597	1.897	2.197	2.497	2.097	2.397	2.697	2.997	2.467	2.767	3.067	3.367
MLP	1.786	2.086	2.386	2.686	2.286	2.586	2.886	3.186	2.656	2.956	3.256	3.556
MLR	1.876	2.176	2.476	2.776	2.376	2.676	2.976	3.276	2.746	3.046	3.346	3.646
RF	2.020	2.320	2.620	2.920	2.520	2.820	3.120	3.420	2.890	3.190	3.490	3.790
MARS	2.098	2.398	2.698	2.998	2.598	2.898	3.198	3.498	2.968	3.268	3.568	3.868

Table 3: Evaluation of the performance of CNN-GRU vs the Benchmark models at Menindee stations

*Models*	*Week 1 (ET₀₇)*		*Week 2 (ET₁₄)*		*Week 3 (ET₂₁)*		*Week 4 (ET₂₈)*
*Models*	r	RMSE	r	RMSE	r	RMSE	r	RMSE
CNN-GRU	0.996	0.126	0.915	0.929	0.905	0.962	0.903	0.982
CNN-LSTM	0.993	0.231	0.912	0.974	0.902	0.974	0.902	0.991
GRU	0.989	0.263	0.917	0.914	0.898	0.995	0.896	1.025
LSTM	0.989	0.236	0.913	0.929	0.904	1.001	0.896	1.007
RNN	0.984	0.280	0.912	0.950	0.904	0.992	0.896	1.116
DTree	0.985	0.407	0.832	1.331	0.767	1.544	0.800	1.48
MLP	0.967	0.599	0.896	1.086	0.899	1.063	0.891	1.10
MLR	0.984	0.262	0.913	0.945	0.894	1.042	0.893	1.051
RF	0.924	0.875	0.882	1.130	0.884	1.09	0.883	1.081
MARS	0.994	0.247	0.911	0.951	0.886	1.054	0.878	1.115

Table 4: Evaluation of the performance of CNN-GRU vs the Benchmark models at Fairfield stations

*Models*	*Week 1* *(ET₀₇)*		*Week 2 (ET₁₄)*		*Week 3 (ET₂₁)*		*Week 4 (ET₂₈)*
*Models*	r	RMSE	r	RMSE	r	RMSE	r	RMSE
CNN-GRU	0.993	0.159	0.912	0.962	0.902	0.995	0.900	1.015
CNN-LSTM	0.990	0.264	0.909	1.007	0.899	1.007	0.899	1.024
GRU	0.986	0.296	0.914	0.947	0.895	1.028	0.893	1.058
LSTM	0.986	0.269	0.910	0.962	0.901	1.034	0.893	1.040
RNN	0.981	0.313	0.909	0.983	0.901	1.025	0.893	1.149
DTree	0.982	0.440	0.829	1.364	0.764	1.577	0.797	1.513
MLP	0.964	0.632	0.893	1.119	0.896	1.096	0.888	1.133
MLR	0.981	0.295	0.910	0.978	0.891	1.075	0.890	1.084
RF	0.921	0.908	0.879	1.163	0.881	1.123	0.880	1.114
MARS	0.987	0.380	0.908	0.984	0.883	1.087	0.775	1.148

Table 5: Evaluation of the performance of CNN-GRU vs. the Benchmark models at Fairfield stations

*Models*	*Week 1* *(ET₀₇)*		*Week 2 (ET₁₄)*		*Week 3 (ET₂₁)*		*Week 4 (ET₂₈)*
*Models*	r	RMSE	r	RMSE	r	RMSE	r	RMSE
CNN-GRU	0.991	0.170	0.910	0.973	0.900	1.006	0.899	1.026
CNN-LSTM	0.988	0.275	0.907	1.018	0.897	1.018	0.898	1.035
GRU	0.984	0.307	0.912	0.958	0.883	1.039	0.892	1.069
LSTM	0.984	0.280	0.908	0.973	0.899	1.045	0.892	1.051
RNN	0.979	0.324	0.907	0.994	0.899	1.036	0.892	1.160
DTree	0.980	0.451	0.827	1.375	0.762	1.588	0.796	1.524
MLP	0.962	0.643	0.891	1.130	0.894	1.107	0.887	1.144
MLR	0.979	0.306	0.908	0.989	0.889	1.086	0.889	1.095
RF	0.919	0.919	0.877	1.174	0.879	1.134	0.879	1.125
MARS	0.987	0.291	0.906	0.995	0.881	1.098	0.874	1.159

Download PDF

Editorial decision: Major revisions
18 May, 2021
Reviews received at journal
16 Apr, 2021
Editor assigned by journal
15 Apr, 2021
Reviewers invited by journal
15 Apr, 2021
First submitted to journal
13 Apr, 2021

You are reading this latest preprint version

Hybrid Deep Learning for Week-Ahead Evapotranspiration Forecasting

Status:

Version 1

Abstract

Figures

Full Text

Tables

Status:

Version 1