Short-term solar radiation forecasting using hybrid deep residual learning and gated LSTM recurrent network with differential covariance matrix adaptation evolution strategy
Introduction
Solar power is one of the most abundant, accessible, and infinitely renewable energy sources yielded when energy from sunlight is transformed into electricity. Solar power is considered a top alternative source to fossil fuels with a high potential to meet global energy demands in the near future [1]. In solar energy technologies, the development of an accurate prediction short-term or long-term (day ahead) [2] solar radiation model plays a fundamental role in enhancing the scheduling and controlling the performance of photovoltaic power plants; having a reliable and robust plan for managing connection to smart grids [3]; and improving the gain margin of the energy suppliers in these markets.
However, predicting solar energy is challenging due to solar radiation’s intermittent and chaotic nature and atmospheric situations that are naturally ungovernable (i.e., clouds, shadows, the vapour of water, ice, air pollution or aerosols in the atmosphere) [4]. Another primary motivation for predicting solar radiation is that installing and maintaining solar radiation measurement devices is highly costly. This makes installing such instruments in every meteorological station financially challenging, especially in developing countries. As an example, there were around 1800 meteorological stations in Turkey in 2020; however, just 7% of them were equipped to register solar radiation data [5]. From this perspective, various technical models have been proposed to forecast solar radiation. The empirical model is one of the popular prediction models established on mathematical procedures. Its benefits include fast and straightforward calculations and is helpful techniques for predicting long-term (monthly or weekly) solar radiation data [6]. However, empirical models cannot accurately predict short-term solar radiation data due to changeable parameters in weather conditions such as cloud cover, rainy days, etc. Furthermore, extracting the intricate and nonlinear associations found in the dependent and independent variables is challenging for empirical models, particularly in humid subtropical climate areas when the weather is rainy with heavy clouds cover [7].
Previous solar energy research studies have proposed considerable number of data-driven techniques for short-term and long-term forecasting. These techniques fall into three main areas, physical techniques, statistical analysis, and machine-deep learning methods. In the physical forecasting models, the atmosphere’s dynamic motion and physical conditions are characterised using a set of mathematical formulas. The performance of physical models relies heavily on the quality and quantity of meteorological variables and astronomical dates (e.g., solar time and earth declination angle) [8]. The statistical approaches using statistical analysis of the various intake features for solar radiation prediction have been applied, including the auto-regressive models (AR), auto-regressive integrated moving average (ARIMA) [9], exponential smoothing, Markov Chain model [10] and Gaussian process [11]. Most of them show acceptable accuracy for predicting the ground solar radiation and cloud motion on different time horizons up to hours ahead. In the last decades, the application of artificial intelligence (AI)-based approaches has considerably developed in solar engineering fields [12]. Previous analyses represent that the AI-based approaches are able to provide more accurate forecasting of solar radiation results than those of the other models [13] such as supervised and unsupervised artificial neural networks (ANN) [14], deep learning models [15], support vector machines (SVM) [16], etc.
A comparative study [17] was done to clarify which one of the six machine learning models can perform best, including the gradient boosting tree (GBT), multi-layer perceptron (MLP), standard ANFIS, subtractive and fuzzy c-means clustering ANFIS, classification and regression tree (CART), and multivariate adaptive regression spline (MARS) to forecast solar irradiation in two sites. The Ref. [17] recommended applying the GBT model as a robust and reliable tool for predicting solar radiation.
A popular sequential deep learning model called long short-term memory (LSTM) is one of the most successful tools in handling the dependency between successive time series data with short-term intervals. One considerable early study was done by Qing and Niu [18] investigated hourly solar radiation forecasting using LSTM. The proposed LSTM [18] was more precise than BPNN in RMSE.
In short-term solar radiation forecasting, recurrent neural networks (RNN) promise high accuracy and robustness in relation to, long short-term memory (LSTM), bidirectional LSTM, and Gated Recurrent Units (GRU). However, initialising the hyper-parameters of RNNs is challenging due to the complex and non-linear relationships between the setting parameters and the topology and nature of the time-series data. To address these problems, Peng et al. [19] developed a hybrid deep learning model combination of BiLSTM, an adaptive ensemble decomposition method (CEEMDAN), and a sine cosine meta-heuristic algorithm (SCA) for predicting hourly stochastic historical time series solar radiation data. The comparative modelling results suggested that the proposed hybrid model [19] could conquer seven other machine learning models.
Nevertheless, increasing the time horizon for forecasting solar radiation is challenging [20] for AI-based methods because of decreased auto-correlation among the time series samples. One preliminary study in long-term global solar radiation by Jiang [21] applied traditional neural networks (feed-forward back-propagation) and compared them with different empirical regression methods. The findings [21] confirmed the superiority and high ability in generalising ANN models demonstrate in solar radiation forecasting. Multilayer perception (MLP) is one of the most popular and classic machine-learning techniques and has been applied in several studies in forecasting solar radiance [22]. In early work [23], Rodriguez et al. applied a combination of five Multilayer perceptron feed-forward neural networks (an ensemble model) developed by a Monte Carlo simulation to forecast global solar radiation, and the overall validation error and accuracy were considerable. However, the drawbacks of fully connected networks (e.g. network overfitting) were not considered in [23]. To deal with the long-term solar forecasting challenges, Kisi [24] analysed and compared the three AI-based methods: fuzzy genetic, ANN, and neuro-fuzzy models for estimating monthly solar radiations from the Mediterranean areas. The modelling results indicated that the fuzzy genetic model could perform better than the other two models. In another study [25] examining monthly solar radiation prediction, a traditional ANN and adaptive neuro-fuzzy inference system (ANFIS) were applied. The prediction results illustrated that ANFIS mostly outperformed other ANNs. However, the study did not discuss the importance of hyper-parameters initialisation [25].
One initial effort to apply deep learning models such as convolutional neural networks (CNN), was proposed by Kaba et al. [26] in order to estimate daily global solar radiation. Although the technical details of the CNN model are not clear, the estimation results show that the CNN model can be an appropriate alternative approach in long-term solar radiation forecasting.
To tackle the issues raised by traditional neural networks training, such as the exhaustive learning process, insufficient parameter preference, and the need for a large number of training samples, a new model of neural network called a Deep Belief Networks (DBN) [27] was proposed in 2007. A combined DBN [28] with a clustering idea was used to develop an accurate daily solar energy forecasting model based on 30 sites located in China. The DBN method [28] acquired more reasonable accuracy from the outcomes than empirical ML methods. Tuning the hyper-parameters of deep learning models is essential; nevertheless, it is frequently challenging. Meta-heuristic algorithms have been applied in order to optimise the hyper-parameters that lead to improving the average performance of the models. Wang et al. [29] proposed a primary hybrid solar radiation forecasting models consisting of an Extreme Learning Machine (ELM) and Cuckoo Search (CS). In order to reduce the computational runtime, the Ref. [29] applied a combination of Multiresponse Sparse Regression (MSR) and leave-one-out cross-validation (LOO-CV) to determine the priority of neurons and remove the lowest ones in Feed Forward Neural Networks. The CS played the role of weight coefficients optimiser.
Despite comprehensive studies in the last years, at least three research gaps remain in designing techniques/models for short- and long-term solar radiation prediction as follows:
- 1.
One of the most significant factors in improving the performance of prediction models hyper-parameters tuning.
- 2.
In most case studies, optimising the architecture of deep learning models using various recurrent neural networks did not consider substantially.
- 3.
The low performance of solar radiation predictors is due to an insufficient decomposition setting .
This study proposes a novel hybrid residual deep learning model for forecasting solar radiation one hour ahead based on real meteorological and Shortwave Radiation (SRAD1) data from an installed offshore buoy station located in Lake Michigan, Chicago, United States and supported by the National Data Buoy Centre (NDBC). In order to clean the data and improve accuracy, an outlier detection method (autoencoder) was applied. A hyper-parameter optimiser is also proposed to reinforce the model’s performance. The foremost contributions of this study are summed as follows:
- 1.
A novel hybrid solar radiation forecasting model is proposed composed of recurrent neural networks (GRU, LSTM and BiLSTM), and a convolutional ResNet50 model (deep residual learning) with adaptive decomposition technique and effective auto-tuner (ADCMA-ResNet50-GRU-2LSTM).
- 2.
An adaptive multivariate empirical mode decomposition (MEMD) algorithm is proposed to decompose solar radiation time-series data with a high level of nonlinearity and non-stationarity into intrinsic mode functions (IMFs) with minimum entropy using an evolutionary Nelder–Mead simplex search algorithm.
- 3.
In order to deal with the shortcomings of the hyper-parameters tuning initialisation, an effective and smart hyper-parameters tuner, adaptive differential covariance matrix evolutionary algorithm (ADCMA), was developed to improve prediction accuracy and reduce modelling bias.
- 4.
A comprehensive comparative framework was also designed to evaluate the performance of various canonical and hybrid machine learning and deep learning models with regard to developing an accurate and reliable solar radiation forecasting model.
The principle sections of this article are organised as follows. The technical details of the involved methods are exemplified in the next Section 2. The following presents the case study’s attributes and their statistical analysis of the dataset used in this study in Section 3. In order to develop a systematic comparison framework for the short-term solar radiation forecast, various models are evaluated and compared with the proposed model in Section 4. Eventually, in Section 5, the acquired results of this investigation and future research plans are outlined. All the acronyms and symbols utilised in this study can be seen in Table 1, Table 2.
Section snippets
Time-domain signal decomposition
Several decomposition techniques can be used to extract the primary characteristics of complex and nonlinear time-series data. They involve a robust statistical approach that disintegrates a nonlinear signal down into some elements based on a directional, periodical and stochastic element. The different applications of these features can forecast, predict or infer unseen data [30]. The most popular time-series decomposition methods are variational mode decomposition (VMD) [31], multivariate VMD
Case study
In this study, we considered the collected real dataset consisting of a combination of solar radiation and standard meteorological data collected from Station OKSI2 - Oak St., Chicago, IL (National Data Buoy Center), from January 2014 to June 2022. The time resolution of data collection was one hour. Fig. 7 shows the geographical location of the station with the online wind speed and air temperature. As the solar radiation data includes many zero values related to nights (can be indicated in
Backbone architecture of recurrent deep learning model
In the first step of this study, nine LSTM models with various features (with one, two, three, and four inputs of SRAD, wind speed, wind direction and air temperature) were proposed and compared to find the best model for the prediction accuracy of solar radiation. A schematic diagram of nine LSTM forecasting models with a representation of the inputs can be seen in Fig. 10. Moreover, We can see the application of LSTM in Fig. 10 as a feature selection technique involves training an LSTM model,
Conclusions
Forecasting short-term solar radiation is challenging because of the intermittent, chaotic nature of solar radiation and atmospheric situations. This article proposed a new hybrid deep learning framework to predict short-term (one-hour ahead) solar irradiance. This framework was used for real data acquired from the National Data Buoy Center, Station OKSI2 - Oak St., Chicago. A detailed pre-processing analysis was applied to detect and clean anomalies, and then the data were normalised. In order
CRediT authorship contribution statement
Mehdi Neshat: Conceptualization, Methodology, Software, Validation, Data curation, Writing – original draft, Writing – review & editing, Visualization. Meysam Majidi Nezhad: Investigation, Conceptualization, Resources, Data curation, Writing – original draft. Seyedali Mirjalili: Investigation, Conceptualization, Supervision, Writing – review & editing. Davide Astiaso Garcia: Supervision, Writing – review & editing. Erik Dahlquist: Supervision, Writing – review & editing. Amir H. Gandomi:
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (67)
Solar energy, governance and CO2 emissions
Renew Energy
(2022)- et al.
Improved complete ensemble empirical mode decomposition with adaptive noise deep residual model for short-term multi-step solar radiation prediction
Renew Energy
(2022) - et al.
Prediction of daily global solar radiation using different machine learning algorithms: Evaluation and comparison
Renew Sustain Energy Rev
(2021) - et al.
The development of empirical models for estimating global solar radiation on horizontal surface: A case study
Renew Sustain Energy Rev
(2018) - et al.
New combined models for estimating daily global solar radiation based on sunshine duration in humid regions: a case study in South China
Energy Convers Manage
(2018) - et al.
Ultra-short-term exogenous forecasting of photovoltaic power production using genetically optimized non-linear auto-regressive recurrent neural networks
Renew Energy
(2021) - et al.
A benchmarking of machine learning techniques for solar radiation forecasting in an insular context
Sol Energy
(2015) - et al.
Taxonomy research of artificial intelligence for deterministic solar power forecasting
Energy Convers Manage
(2020) - et al.
Comparison of support vector machine and copula-based nonlinear quantile regression for estimating the daily diffuse solar radiation: A case study in China
Renew Energy
(2020) - et al.
Deep learning models for solar irradiance forecasting: A comprehensive review
J Clean Prod
(2021)
Application of support vector machine models for forecasting solar and wind energy resources: A review
J Clean Prod
A comparative study of several machine learning based non-linear regression methods in estimating solar radiation: Case studies of the USA and Turkey regions
Energy
Hourly day-ahead solar irradiance prediction using weather forecasts by LSTM
Energy
An integrated framework of bi-directional long-short term memory (BiLSTM) based on sine cosine algorithm for hourly solar radiation forecasting
Energy
Computation of monthly mean daily global solar radiation in China using artificial neural networks and comparison with other empirical models
Energy
Uncertainties in global radiation time series forecasting using machine learning: The multilayer perceptron case
Energy
An artificial neural network ensemble model for estimating global solar radiation from meteosat satellite images
Energy
Modeling solar radiation of mediterranean region in Turkey by using fuzzy genetic approach
Energy
Artificial neural networks and adaptive neuro-fuzzy inference systems approaches to forecast the meteorological data for HVAC: The case of cities for Turkey
Energy
Estimation of daily global solar radiation using deep learning model
Energy
Application of functional deep belief network for estimating daily global solar radiation: A case study in China
Energy
Forecasting solar radiation using an optimized hybrid model by cuckoo search algorithm
Energy
A review and discussion of decomposition-based hybrid models for wind energy forecasting applications
Appl Energy
A method for the generation of typical meteorological year data using ensemble empirical mode decomposition for different climates of China and performance comparison analysis
Energy
Multivariate empirical mode decomposition based hybrid model for day-ahead peak load forecasting
Energy
Predictability analysis of absence seizures with permutation entropy
Epilepsy Res
Causal convolutional gated recurrent unit network with multiple decomposition methods for short-term wind speed forecasting
Energy Convers Manage
Review of differential evolution population size
Swarm Evol Comput
Generalized normal distribution optimization and its applications in parameter extraction of photovoltaic models
Energy Convers Manage
The arithmetic optimization algorithm
Comput Methods Appl Mech Engrg
A differential covariance matrix adaptation evolutionary algorithm for real parameter optimization
Inform Sci
A hybrid VMD-LSTM/GRU model to predict non-stationary and irregular waves on the east coast of China
Ocean Eng
A deep learning-based evolutionary model for short-term wind speed forecasting: A case study of the lillgrund offshore wind farm
Energy Convers Manage
Cited by (18)
A short-term wave energy forecasting model using two-layer decomposition and LSTM-attention
2024, Ocean EngineeringAn innovative end-to-end PINN-based solution for rapidly simulating homogeneous heat flow problems: An adaptive universal physics-guided auto-solver
2024, Case Studies in Thermal EngineeringRead-First LSTM model: A new variant of long short term memory neural network for predicting solar radiation data
2024, Energy Conversion and Management