Learning Long Short-term Memory Research Articles

A Hybrid Deep Learning Model to Estimate the Future Electricity Demand of Sustainable Cities

Rapid population growth, economic growth, and technological developments in recent years have led to a significant increase in electricity consumption. Therefore, the estimation of electrical energy demand is crucial for the planning of electricity generation and consumption in cities. This study proposes a hybrid deep learning model that combines convolutional neural network (CNN) and long short-term memory (LSTM) techniques, both of which are deep learning techniques, to estimate electrical load demand. A hybrid deep learning model and LSTM model were applied to a dataset containing hourly electricity consumption and meteorological information of a city in Türkiye from 2017 to 2021. The results were evaluated using mean absolute percent error (MAPE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) metrics. The proposed CNN-LSTM hybrid model was compared to the LSTM model, with lower MAPE, MAE, and RMSE values. Furthermore, the CNN-LSTM model exhibited superior prediction performance with an R2 value of 0.8599 compared to the LSTM model with an R2 value of 0.8086. These results demonstrate the effectiveness of the proposed deep learning model in accurately estimating future electrical load demand to plan electricity generation for sustainable cities.

Ensemble empirical mode decomposition based deep learning models for forecasting river flow time series

River flow forecasting is important for flood prediction and effective utilization of water resources. This study proposed a comprehensive methodology that simultaneously enables the creation of multiple datasets and a comparison of multiple learning models for forecasting river flows. First, the data characteristics of the river flow time series were assessed based on its nature and pattern characteristics. Second, data decomposition was made for simplifying the complex data into decomposed modes. Ensemble Empirical Mode Decomposition (EEMD) was used for this purpose that provides intrinsic mode functions (IMFs). In addition, as benchmark, traditional time series decomposition was also applied over the complex data. Third, based on their data characteristics, the IMFs were combined into different components. These components along with the traditional time series decomposition create multiple datasets. Then, the components from the different datasets were forecasted using deep learning models, mainly BPNN (Back-Propagation Neural Network), CNN (Convoluted Neural Network) and LSTM (Long-Short Term Memory). Then, ensemble prediction was used to get the final output. For verification and benchmark purpose, we have also used statistical models like SARIMA over the datasets. A comprehensive evaluation and model comparisons were made using Diebold-Mariano (DM) test, k-fold cross validation and with the results of Prophet. In addition, the temporal convolution network (TCN) and the gated recurrent unit (GRU) were also modelled. Finally, all the DL modelling was performed for two data situations: with missing data deleted and missing data imputed. An empirical case study was done considering the average of daily time series data of Neeleswaram Hydrological Observation (HO) station over the Periyar river of South India for the period 2017–2021. The empirical assessment was made using several tests for all the different datasets and different learning models for further insights. Our results indicate that the deep learning-based models offer better predictions for complex time series in comparison to traditional statistical models. The DL models applied on the original time series data provided robust predictions with good performance such as low RMSE and high correlation values. However, the DL models’ performances are not enhanced through integration with time series decomposition and EEMD. Furthermore, the empirical results and the k-fold cross validation-based test indicate that all the DL models show equivalent performances. Interestingly, the DL models show superior performance over Prophet.

Slope multi-step excavation displacement prediction surrogate model based on a long short-term memory neural network: for small sample data and multi-feature multi-task learning

ABSTRACT Machine learning-based surrogate models have become the preferred approach for large-scale and frequent simulation tasks due to its significant improvement in computational efficiency. In order to overcome the potential effects of learning with small sample data and the challenges of multi-feature multi-task learning, we developed a novel deep learning long short-term memory (LSTM) model. Taking slope excavation displacement prediction as a case study, we employed the Latin hypercube sampling method to generate a synthetic dataset for training LSTM and other mainstream models. Experimental results demonstrate that the model's prediction accuracy decreases with a reduction in sample size, while support vector regression (SVR), back propagation neural network (BPNN), LSTM and Gaussian process regression (GPR) demonstrate a stronger resistance. It is feasible to utilise excavation features as model inputs to establish a unified multi-step excavation model, but the accuracy of the SVR model decreased by 32.5% after supplementing excavation features. Even when the sample size is less than 50, both LSTM and GPR exhibit excellent performance, achieving model R-squared and RMSE surpassing 0.99 and 0.07 mm. However, when addressing multi-output learning tasks, LSTM stands out as the optimal choice. This study will assist researchers or engineers in swiftly selecting appropriate surrogate models.

Satellite-based soybean yield prediction in Argentina: A comparison between panel regression and deep learning methods

The accurate prediction of soybean yield is vital for global food market stabilization and food security. Recent advancements in remote sensing technology have significantly amplified interest in leveraging satellite-based methods for predicting crop yield. These methods offer in-season yield estimates. By utilizing this timely information, decision-makers can formulate strategic, well-informed choices that preemptively mitigate potential food price hikes, ultimately bolstering food security. While simple regression models have been widely utilized for satellite-based yield prediction, researchers have recently begun to explore the use of deep learning algorithms. This study compares the performance of panel regression and deep learning models for in-season soybean yield prediction at the Department (county-equivalent) level in Argentina. Data sources include the latest soybean land use products and MODIS bi-weekly vegetation index products. Results indicate that deep learning models significantly outperform panel regression. Deep learning Long Short-Term Memory (LSTM) models, which incorporate attention mechanism and a series of peak NDVI images, generate more accurate and time-sensitive predictions. Among competing LSTM models, the one with attention mechanism applied to the entire growing season's NDVI data yields the highest prediction accuracy, with a Root Mean Square Error (RMSE) of 505.78 kg/ha and Normalized Root Mean Square Error (NRMSE) of 0.0726. The LSTM model with attention on the three highest NDVI images attains a satisfactory prediction accuracy (RMSE = 627.28 kg/ha, NRMSE = 0.089) six weeks prior to harvest. This study presents a robust model for predicting crop yields, promoting sustainable production of soybeans and facilitating knowledgeable choices among farmers and policymakers.

Cortical signals analysis to recognize intralimb mobility using modified RNN and various EEG quantities

Electroencephalogram (EEG) signals are critical in interpreting sensorimotor activities for predicting body movements. However, their efficacy in identifying intralimb movements, such as the dorsiflexion and plantar flexion of the foot, remains suboptimal. This study aims to explore whether various EEG signal quantities can effectively recognize intralimb movements to facilitate the development of Brain-Computer Interface (BCI) devices for foot rehabilitation. This research involved twenty-two healthy, right-handed participants. EEG data were collected using 21 electrodes positioned over the motor cortex, while two electromyography (EMG) electrodes recorded the onset of ankle joint movements. The study focused on analyzing slow cortical potential (SCP) and sensorimotor rhythms (SMR) in alpha and beta bands from the EEG. Five key features—fourth-order Autoregressive feature, variance, waveform length, standard deviation, and permutation entropy—were extracted. A modified Recurrent Neural Network (RNN) including Long Short-term Memory (LSTM) and Gated Recurrent Unit (GRU) algorithms was developed for movement recognition. These were compared against conventional machine learning algorithms, including nonlinear Support Vector Machine (SVM) and k Nearest Neighbourhood (kNN) classifiers. The performance of the proposed models was assessed using two data schemes: within-subject and across-subjects. The findings demonstrated that the GRU and LSTM models significantly outperformed traditional machine learning algorithms in recognizing different EEG signal quantities for intralimb movement. The study indicates that deep learning models, particularly GRU and LSTM, hold superior potential over standard machine learning techniques in identifying intralimb movements using EEG signals. Where the accuracies of LSTM for within and across subjects were 98.87 ± 1.80 % and 87.38 ± 0.86 % respectively. Whereas the accuracy of GRU within and across subjects were 99.18 ± 1.28 % and 86.44 ± 0.69 % respectively. This advancement could significantly benefit the development of BCI devices aimed at foot rehabilitation, suggesting a new avenue for enhancing physical therapy outcomes.

Deep Learning Models for Bitcoin Prediction Using Hybrid Approaches with Gradient-Specific Optimization

Since cryptocurrencies are among the most extensively traded financial instruments globally, predicting their price has become a crucial topic for investors. Our dataset, which includes fluctuations in Bitcoin’s hourly prices from 15 May 2018 to 19 January 2024, was gathered from Crypto Data Download. It is made up of over 50,000 hourly data points that provide a detailed view of the price behavior of Bitcoin over a five-year period. In this study, we used potent algorithms, including gradient descent, attention mechanisms, long short-term memory (LSTM), and artificial neural networks (ANNs). Furthermore, to estimate the price of Bitcoin, we first merged two deep learning algorithms, LSTM and attention mechanisms, and then combined LSTM-Attention with gradient-specific optimization to increase our model’s performance. Then we integrated ANN-LSTM and included gradient-specific optimization for the same reason. Our results show that the hybrid model with gradient-specific optimization can be used to anticipate Bitcoin values with better accuracy. Indeed, the hybrid model combines the best features of both approaches, and gradient-specific optimization improves predictive performance through frequent analysis of pricing data changes.

Cryptocurrency assets valuation prediction based on LSTM, neural network, and deep learning hybrid model

Cryptocurrency, a digital currency managed by decentralized networks, has gained immense popularity since the inception of Bitcoin. These digital assets, often characterized by extreme price volatility, have generated substantial interest from investors. Traditional financial models struggle to account for the unique dynamics and complexities of cryptocurrencies, prompting the adoption of deep learning techniques. This study investigates the use of Long Short-Term Memory (LSTM), Neural Networks, and Deep Learning (CNN) in predicting cryptocurrency prices. These deep learning models leverage various data sources, such as technical indicators and sentiment analysis, to gain a comprehensive understanding of cryptocurrency markets. The research evaluates the performance of these models using Root Mean Squared Error (RMSE) as the primary metric. The results demonstrate that the hybrid model, combining LSTM, Neural Networks, and Deep Learning, exhibits the highest predictive accuracy across multiple cryptocurrencies, including Bitcoin (BTC), Ethereum (ETH), and Binance Coin (BNB). However, challenges persist, such as model adaptability to unforeseen market events and data noise. Future developments may involve incorporating external factors and interdisciplinary collaboration to create more holistic valuation models. Despite these challenges, the study underscores the potential of hybrid deep learning models in enhancing cryptocurrency valuation accuracy and their relevance in risk management strategies for investors and traders.

Shape Recognition and Corner Points Detection in 2D Drawings Using a Machine Learning Long Short-Term Memory (LSTM) Approach

Creating a 2D geometry model from an image poses challenges for CAD users due to factors such as noise, segmentation difficulties, complex geometric structures, scale and perspective variations, and the need for CAD system compatibility. In this paper, we propose a novel deep learning approach utilizing Long-Short Term Memory (LSTM) to address these challenges. Our approach decomposes the shapes in the images into line and curve segments and accurately locates their intersection points. To enhance the model’s performance, we introduce two distinct types of features (angle and curvature features) and optimize the model through hyperparameter tuning. The resulting model exhibits robustness against noise, varying image sizes, and can effectively locate different types of intersection points. To evaluate the proposed model, we have developed a Python-based software and conducted experiments on a dataset comprising of 200 shapes with seven different resolutions. Comparative analysis against a state-of-the- art method (TCVD) from the literature demonstrates that our approach achieves higher accuracy in terms of line, curve, and intersection point detection.

Ship SINS/CNS Integrated Navigation Aided by LSTM Attitude Forecast

Under the strong interference of sky background noise, the reliability of celestial navigation system (CNS) measurement will drop sharply, which leads to performance deterioration for ships’ strapdown inertial navigation system (SINS)/CNS integrated navigation. To solve this problem, a long short-term memory (LSTM) model is trained to forecast a ship’s attitude to detect the attitude provided by the CNS, and the LSTM forecasted attitude can also be used as a backup in case of CNS failure. First, the SINS/CNS integrated model is derived based on an attitude solution of the CNS, which provides more favorable feature data for LSTM learning. Then, the key techniques of LSTM modeling such as dataset construction, LSTM coding method, hyperparameter optimization and training strategy are described in detail. Finally, an experiment is conducted to evaluate the actual performance of the investigated methods. The results show that the LSTM model can accurately forecast a ship’s attitude: the horizon reference error is less than 0.5′ and the yaw error is less than 0.6′, which can provide reliable reference attitude for the SINS when the CNS is invalid.

Forecasting the high‐frequency volatility based on the LSTM‐HIT model

AbstractVolatility forecasting from high‐frequency data plays a crucial role in many financial fields, such as risk management, option pricing, and portfolio management. Many existing statistical models could better describe and forecast the characteristics of volatility, whereas they do not simultaneously account for the long‐term memory of volatility, the nonlinear characteristics of high‐frequency data, and technical index information during the modeling phase. The purpose of this paper is to use the prediction advantage of deep learning long short‐term memory (LSTM) model to predict the volatility fusing three classes of information, that is, high frequency realized volatility (H), technical indicators (I), and the parameters of generalized autoregression conditional heteroskedasticity(GARCH), heterogeneous autoregressive (HAR), and c, resulting in a novel LSTM‐HIT model to forecast realized volatility. We employ the extreme value theory (EVT) of a semiparametric method to estimate the quantile of standardized return and construct the LSTM‐HIT‐EVT model to forecast the value at risk (VaR). Empirical results show that the LSTM‐HIT model provides the most accurate volatility forecast among the various considered models and that the LSTM‐HIT‐EVT model yields forecasts more accurate than other VaR models.

Comparative Analysis of Machine Learning and Deep Learning Models for Bitcoin Price Prediction

This research endeavors to forecast Bitcoin prices by employing a suite of machine learning and deep learning models. Five distinct models were employed: Random Forest, Linear Regression, Support Vector Machine (SVM), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU), each evaluated based on their R-squared scores. Notably, the models showcased diverse performances, with the ensemble learning approach of Random Forest exhibiting near-perfect accuracy, closely followed by GRU and SVM. The deep learning architectures, LSTM and GRU, demonstrated remarkable predictive capabilities, showcasing their adeptness in capturing intricate temporal patterns within the cryptocurrency price data. This study sheds light on the comparative performance of these models, emphasizing their strengths and limitations in predicting Bitcoin prices.

A wind power forecasting model based on data decomposition and cross-attention mechanism with cosine similarity

In the context of societal advancement, escalating energy demands underscore the critical significance of exploring renewable resources. The accurate prediction of wind power is crucial for maximizing the utilization of wind energy and ensuring the smooth operation of wind farms. Currently, the majority of models are built upon architectures such as Transformer and Long-Short Term Memory (LSTM). However, Transformer-type models are not well-suited for time series forecasting, and LSTM often focuses on historical sequence information, exhibiting limited capability in capturing the relevant information between weather factors and power. Recognizing this inadequacy, this study proposes a novel model—a linear network architecture featuring a cosine-related cross-attention mechanism. The signal decomposition is carried out utilizing variational mode decomposition, with the determination of the optimal decomposition count guided by sample entropy analysis, named Improved Variational Mode Decomposition (IVMD). Subsequently, the conventional self-attention mechanism is replaced with a cross-attention mechanism. Contemporary mainstream models are primarily based on improved versions of core algorithms such as LSTM, Transformer, and Extreme Learning Machines. Experimental results show significant performance enhancements compared to these models across different timescales. In short-term, medium-term, and long-term forecasts, accuracy improves by more than 40 %, 20 %, and 5 %, respectively.

Joint torque prediction of industrial robots based on PSO-LSTM deep learning

PurposeBecause of the key role of joint torque in industrial robots (IRs) motion performance control and energy consumption calculation and efficiency optimization, the purpose of this paper is to propose a deep learning torque prediction method based on long short-term memory (LSTM) recurrent neural networks optimized by particle swarm optimization (PSO), which can accurately predict the the joint torque.Design/methodology/approachThe proposed model optimized the LSTM with PSO algorithm to accurately predict the IRs joint torque. The authors design an excitation trajectory for ABB 1600–10/145 experimental robot and collect its relative dynamic data. The LSTM model was trained with the experimental data, and PSO was used to find optimal number of LSTM nodes and learning rate, then a torque prediction model is established based on PSO-LSTM deep learning method. The novel model is used to predict the robot’s six joint torque and the root mean error squares of the predicted data together with least squares (LS) method were comparably studied.FindingsThe predicted joint torque value by PSO-LSTM deep learning approach is highly overlapped with those from real experiment robot, and the error is quite small. The average square error between the predicted joint torque data and experiment data is 2.31 N.m smaller than that with the LS method. The accuracy of the novel PSO-LSTM learning method for joint torque prediction of IR is proved.Originality/valuePSO and LSTM model are deeply integrated for the first time to predict the joint torque of IR and the prediction accuracy is verified.

Drought index downscaling using AI-based ensemble technique and satellite data

This study introduces and validates an artificial intelligence (AI)–based downscaling method for Standardized Precipitation Indices (SPI) in the northwest of Iran, utilizing PERSSIAN-CDR data and MODIS-derived drought-dependent variables. The correlation between SPI and two drought-dependent variables at a spatial resolution of 0.25° from 2000 to 2015 served as the basis for predicting SPI values at a finer spatial resolution of 0.05° for the period spanning 2016 to 2021. Shallow AI models (Support Vector Regression, Adaptive Neural Fuzzy Inference System, Feedforward Neural Network) and the Long Short-Term Memory (LSTM) deep learning method are employed for downscaling, followed by an ensemble post-processing technique for shallow AI models. Validation against rain gauge data indicates that all methods improve SPI simulation compared to PERSIANN-CDR products. The ensemble technique excels by 20% and 25% in the training and test phases, respectively, achieving the mean Determination Coefficient (DC) score of 0.67 in the validation phase. Results suggest that the deep learning LSTM method is less suitable for limited observed data compared to ensemble techniques. Additionally, the proposed methodology successfully detects approximately 80% of drought conditions. Notably, SPI-6 outperforms other temporal scales. This study advances the understanding of AI-driven downscaling for SPI, emphasizing the efficacy of ensemble approaches and providing valuable insights for regions with limited observational data.

Diagnosis of pathological speech with streamlined features for long short-term memory learning

Background:Pathological speech diagnosis is crucial for identifying and treating various speech disorders. Accurate diagnosis aids in developing targeted intervention strategies, improving patients’ communication abilities, and enhancing their overall quality of life. With the rising incidence of speech-related conditions globally, including oral health, the need for efficient and reliable diagnostic tools has become paramount, emphasizing the significance of advanced research in this field. Methods:This paper introduces novel features for deep learning in the analysis of short voice signals. It proposes the incorporation of time-space and time–frequency features to accurately discern between two distinct groups: Individuals exhibiting normal vocal patterns and those manifesting pathological voice conditions. These advancements aim to enhance the precision and reliability of diagnostic procedures, paving the way for more targeted treatment approaches. Results:Utilizing a publicly available voice database, this study carried out training and validation using long short-term memory (LSTM) networks learning on the combined features, along with a data balancing strategy. The proposed approach yielded promising performance metrics: 90% accuracy, 93% sensitivity, 87% specificity, 88% precision, an F1 score of 0.90, and an area under the receiver operating characteristic curve of 0.96. The results surpassed those obtained by the networks trained using wavelet-time scattering coefficients, as well as several algorithms trained with alternative feature types. Conclusions:The incorporation of time–frequency and time-space features extracted from short segments of voice signals for LSTM learning demonstrates significant promise as an AI tool for the diagnosis of speech pathology. The proposed approach has the potential to enhance the accuracy and allow for real-time pathological speech assessment, thereby facilitating more targeted and effective therapeutic interventions.

Daily flow discharge prediction using integrated methodology based on LSTM models: Case study in Brahmani-Baitarani basin

For flood control, hydropower operation, and agricultural planning, among other applications, flow discharge prediction is a critical first step toward the strong and dependable planning and management of water resources. Floods are destructive natural calamities that destroy human lives and infrastructure across the world. Development of effective flood forecasting and prediction models is critical for minimising deaths and mitigating damages. This study employs hybrid deep learning Long Short Term Memory (LSTM) algorithms like LSTM, Convolution LSTM (Conv-LSTM) and Convolutional Neural Network LSTM (CNN-LSTM) to predict likelihood flood events using daily precipitation, daily temperature and daily relative humidity from two flood-forecasting stations i.e., Champua (Baitarani River, Odisha) and Jarikela (Brahmani River, Odisha) over a 20-year period. The results show that CNN-LSTM performed best followed by Conv-LSTM and LSTM in terms of R2 = 0.98055, 0.96564, and 0.93244, RMSE = 19.137, 35.635, and 49.347, MAE = 18.372, 33.766, and 47.058, NSE = 0.971, 0.9517 and 0.9257 respectively. The findings support the claim that machine learning models and algorithms, in particular CNN-LSTM model, can be applied to flood forecasting with high accuracy, thereby enhancing water and hazard management.

Machine learning in electron beam lithography to boost photoresist formulation design for high-resolution patterning.

The reduction of the critical dimension (CD) usually improves the resolution of patterns and performance of chips. In chip manufacturing, electron beam lithography (EBL) is a promising technology for preparing sub-10 nm patterns, and its imaging resolution is primarily determined by the photoresist formulation. However, the smaller CDs are mainly achieved by optimizing process conditions, and little attention has been paid to the photoresist formulation optimization. Screening suitable photoresist formulations remains a significant challenge due to the considerable time and high cost. Herein, we report a formulation optimization technique of a metal oxide nanoparticle photoresist that combines EBL experiments with a machine learning long short-term memory (LSTM) network. Using the LSTM network, a CD photoresist evaluation model is established. Leveraging the CD model, a photoresist formulation optimizer is developed with a line width of 26 nm. The verification results demonstrate that the CDs predicted by the LSTM network are basically consistent with the EBL experimental results, and the photoresist formulations that meet the CD requirements can be screened. This work opens up a novel perspective to boost photoresist formulation design for high-resolution patterning with artificial intelligence and provides guidance for EBL experiments.

Deep Learning Based Method for Early Warning of Price Risk in Railroad Bulk Freight Transport with Consideration of Time Series of Comprehensive Impact Index

The aim in this study is to enhance the price competitiveness of medium- and long-haul railroad bulk cargoes in the transportation market. Efficiently triggering an early warning for railroad bulk freight price risk is crucial. This paper discusses relevant factors affecting the pricing of railway freight from four aspects, namely the macroeconomy, the transportation market, the freight owner, and the railway enterprise. We construct a comprehensive index system for generating an early warning of railroad bulk freight prices. Utilizing a collected data set of 20 factor indicators that affected prices from 2015 to 2017, we calculate a comprehensive price risk warning index using the integrated entropy weight and the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) method. We propose a risk warning level classification method based on the k-means clustering method to determine five levels of price risk warning. Subsequently, we establish comprehensive index prediction methods for issuing price risk warnings based on the back propagation (BP) neural network and the deep learning long short-term memory (LSTM) network. We verify that the prediction performance of the deep learning LSTM network is significantly better than that of the BP neural network by combining all coal freight data from the Shanghai Railway Bureau for the period 2015 to 2017 as a case study. The proposed method can predict the level of freight price risk more accurately in the next phase, based on the status of the factor indicators, and be used to assist railway departments in developing a more reasonable freight price adjustment strategy to scientifically mitigate freight price risk.

Deep learning integration optimization of electric energy load forecasting and market price based on the ANN–LSTM–transformer method

Introduction: Power load forecasting and market price analysis have become crucial in the context of complex power energy systems and volatile market prices. Deep learning technology has gained significant attention in time series forecasting, and this article aims to enhance the accuracy and reliability of power load and market price predictions by integrating and optimizing deep learning models.Methods: We propose a deep learning framework that combines artificial neural networks (ANNs), long short-term memory (LSTM), and transformer models to address key challenges in electricity load forecasting and market price prediction. We leverage ANNs for their versatility and use LSTM networks for sequence modeling to generate initial predictions. Additionally, we introduce transformer technology and utilize its self-attention mechanism to capture long-distance dependencies within the data, further enhancing the model’s performance.Results: In our experiments, we validate the proposed framework using multiple public datasets. We compare our method with traditional forecasting approaches and a single-model approach. The results demonstrate that our approach outperforms other methods in predicting power load and market prices. This increased accuracy and reliability in forecasting can be of significant value to decision-makers in the energy sector.Discussion: The integration of deep learning models, including ANN, LSTM, and transformer, offers a powerful solution for addressing the challenges in power load and market price prediction. The ability to capture long-distance dependencies using the transformer's self-attention mechanism improves forecasting accuracy. This research contributes to the field of energy and finance by providing a more reliable framework for decision-makers to make informed choices in a complex and dynamic environment.

CSG compressive strength prediction based on LSTM and interpretable machine learning

Abstract As a new type of environmentally friendly building material, cemented sand and gravel (CSG) has advantages distinct from those of concrete. Compressive strength is an important mechanical property of CSG. However, his method of testing is mainly by doing experiments. For this reason, a deep learning algorithm, long short-term memory (LSTM) model, was proposed to predict the compressive strength of CSG by using four input variables, namely cement content, sand rate, water-binder ratio, and fly ash content, with a total of 114 sample data. Three metrics – coefficient (R 2), root mean square error (RMSE), and mean absolute error (MAE) – were used to evaluate the model’s performance, and the predicted results were compared with the traditional machine learning algorithm, namely the random forest (RF) model. Finally, SHapley Additive exPlanations can be combined to explain the contribution degree of each input feature in the machine learning inquiry model to the prediction results. The results show that the prediction accuracy and reliability of LSTM are higher. The LSTM model has R 2 = 0.9940, RMSE = 0.1248, and MAE = 0.0960, while the RF model has R 2 = 0.9147, RMSE = 0.4809, and MAE = 0.4397. The LSTM model can accurately predict CSG compressive strength. Cement and sand rate contribute more to the predicted results than other input characteristics.