Research on a Deeply Integrated Model for Structural Optimization in Coal Spontaneous Combustion Temperature Prediction

Summary

Here, we describe a protocol for predicting the temperature of coal spontaneous combustion using an SSA-optimized CNN-LSTM-Attention framework that automatically optimizes the network structure and parameters, thereby improving accuracy, adaptability, and generalization across heterogeneous datasets and varied mining conditions.

Abstract

Here, we propose a protocol for coal spontaneous combustion temperature prediction based on a Sparrow Search Algorithm (SSA)-optimized convolutional neural networks (CNN)-long short-term memory (LSTM)-Attention framework. This protocol addresses the limitations of fixed network architectures, restricted generalization, and poor transferability commonly encountered in conventional methods. The framework extracts spatial features using CNN and captures temporal dependencies with LSTM networks, while the attention mechanism highlights critical temperature phases and salient features. The SSA jointly optimizes network depth and hyperparameters, enabling dynamic adaptation to varying data complexities across different mining sites and experimental conditions. The protocol consists of data acquisition, feature preprocessing, model construction, parameter optimization, and validation steps. Experimental results demonstrate that the proposed model achieves significantly higher predictive accuracy on homogeneous datasets and maintains robust generalization performance across heterogeneous datasets, making it well-suited for real-time coal mine temperature monitoring and early-warning systems.

Introduction

Coal still plays a dominant role in China's energy structure. However, during storage, transportation, and mining, spontaneous heating can occur, leading to spontaneous combustion. This often causes mine fires, seriously threatening mine safety and workers' lives^1,2,3,4. Hence, accurate prediction of coal mine fire risks and their temperature variations is essential for early warning and disaster mitigation. The prediction methods for coal spontaneous combustion temperature have evolved from early empirical formulas to analysis approaches based on actual monitoring data^5,6,7. Early studies mainly relied on physical models and chemical kinetics equations to model the coal oxidation heat release mechanism. However, owing to the complexity of model parameters and limited applicability, these models face significant limitations in practical use^8,9,10. With the development of intelligent algorithms, methods such as Random Forest (RF)¹¹ and Artificial Neural Networks (ANN)¹² have been gradually introduced to enhance prediction capabilities by constructing nonlinear mappings. In recent years, advances in deep learning have offered new approaches for predicting coal spontaneous combustion temperature. Wang et al.¹³ proposed a detection method based on acoustic temperature measurement technology, analyzing gas emissions under various conditions to establish preliminary thresholds and an early warning system. It developed a refined mathematical model for measuring the temperature of loose coal. Graph Neural Networks (GNN) and Transformer architectures have also been introduced. Pan et al.³ presented a coal spontaneous combustion temperature prediction model based on Graph Convolutional Networks (GCN), which incorporates the interactions among gas indicators to achieve higher prediction accuracy.

However, existing methods still face limitations in hyperparameter selection and model generalization. To improve performance, researchers have introduced intelligent optimization algorithms such as the Sparrow Search Algorithm (SSA) to tune hyperparameters like learning rate and batch size. Wang et al.¹⁴ proposed the SSA-CNN model, which, for the first time, integrates a swarm intelligence algorithm with a convolutional neural network structure. This approach not only improves parameter optimization efficiency but also effectively captures the spatial structural features in coal spontaneous combustion data. Long et al.¹⁵ and Zou et al.¹⁶, respectively, proposed coal spontaneous combustion prediction models based on BO-Light Gradient-Boosting Machine (GBM) and particle swarm optimization-XGBoost (PSO-XGB). Both models improved convergence and accuracy by optimizing the search strategy, providing new approaches for the optimization of coal spontaneous combustion prediction models.

Although existing studies have advanced the intelligent prediction of coal spontaneous combustion temperature, most efforts have been limited to optimizing model parameters without improvements at the network architecture level. Consequently, current models commonly exhibit the following limitations: First, most adopt static structures, where the network architecture (e.g., number of convolutional layers, LSTM layers) and key training parameters (e.g., learning rate, batch size) are manually set or proportionally decayed during initial model construction and remain unchanged throughout training and prediction, lacking the ability to dynamically adapt to data complexity. Second, the models generally lack adaptive mechanisms to adjust for varying conditions such as different time periods, temperature ranges, and gas concentration scales, making it difficult to meet the demands of multi-condition prediction. Third, their generalization and transferability remain inadequate, resulting in unstable and inaccurate predictions across different regional datasets. Although some improvements have been achieved at the parameter level, the network structure itself remains static without joint dynamic optimization of structure and parameters, limiting overall performance gains.

Therefore, this study aims to develop a flexible and high-precision model for coal spontaneous combustion temperature prediction to address the above challenges. Through comparative experiments, the CNN-LSTM-Attention architecture was selected as the base model. To accommodate varying network depth requirements for different coal seams, the traditional static structure is replaced with a "structural optimization" approach. In this approach, the number of convolutional (CNN) and recurrent (LSTM) layers, along with key training parameters, is not fixed but is dynamically adjusted according to the strength of spatial and temporal features in the data.

Protocol

1.Coal spontaneous combustion data collection

1. Assemble the air-type coal programmed heating oxidation system, including the gas delivery system, temperature control unit, and gas analysis instruments. Ensure that all connections are secure, temperature control settings are accurate, and the gas chromatograph and other analytical devices are fully calibrated.
2. Weigh 1000 g of mixed coal sample (keep the original lump form), thoroughly homogenize the sample using the quartering method, and evenly place the coal blocks inside the heating chamber. Maintain consistent thickness and uniform distribution to ensure even heating.
3. Start the air pump and adjust the flow rate using the flow meter to maintain a stable gas flow of 100 mL/min into the heating chamber. Ensure the flow stability within ±2 mL/min.
4. Initiate the heating system according to the preset temperature ramp program, gradually increasing the furnace temperature at a rate of 1.0 °C/min until 200 °C. Use the control software to monitor temperature changes in real time to ensure the heating rate meets the design specifications.
5. Use temperature sensors installed near the coal sample to continuously monitor and record temperature variations, ensuring accurate and uninterrupted data collection.
   NOTE: During the programmed heating experiment, temperature and gas concentration data were automatically recorded every 10 s using calibrated temperature sensors and gas analyzers until the experiment was completed.
6. Transport the gases generated during the heating reaction through the gas delivery system into the gas chromatograph.
   1. Perform automated sampling at predefined intervals (30 s) to measure the concentrations of oxygen, carbon monoxide, and carbon dioxide, and perform GC baseline calibration before each experiment.
   2. Maintain a transfer-line flow of 100 mL/min (±2 mL/min), matched to the chamber outflow and regulated by the mass-flow controller to ensure stable delivery to the gas chromatograph during the 30-s autosampling.
7. Before model training, apply a standardized data preprocessing pipeline to ensure data quality and consistency.
   1. Normalize all features to zero mean and unit variance to eliminate scale discrepancies and facilitate model convergence.
   2. To suppress sensor noise and stabilize temporal trends, apply a moving average filter with a window size of 5 to smooth the time-series data. Impute missing values, if present, using the mean of neighboring observations to maintain dataset completeness.
   3. Furthermore, identify outliers using a z-score threshold of |z| > 3 and exclude them from subsequent analysis to improve model robustness.
      NOTE: The experiment collected data from 83 coal samples, recording the concentrations of gases such as oxygen, carbon monoxide, carbon dioxide, methane, ethane, and ethylene in relation to the coal temperature. A schematic diagram of the air-type coal programmed heating oxidation system is shown in Figure 1.

2. Construction of a CNN-LSTM-attention model based on SSA structural optimization

1. Construction of the CNN-LSTM-attention model
   1. Build the convolutional neural network (CNN) module to automatically extract spatial features from input sequences of temperature and gas concentrations.
      1. Stack 2-4 1D convolution layers (stride = 1, padding = "same", kernel size 3-7). After each layer, apply normalization (BatchNorm or LayerNorm), ReLU or GELU activation, and 0.1-0.2 dropout. Optionally use MaxPooling1D (pool = 2) in the first one or two layers for denoising/downsampling, and avoid pooling in the final layer to preserve temporal resolution.
      2. Use a 1×1 convolution to project the channel dimension to what the downstream LSTM expects. Keep the tensor in a 3D "time × channels" shape (do not flatten) and feed it directly to the LSTM with return_sequences=True; record layer count, channels, kernel size, pooling, and dropout in a config file to enable SSA search and reproducibility.
   2. Treat the final Conv1D output as a time-ordered sequence of feature vectors, preserving the temporal axis without flattening, and feed it directly into the LSTM. If prior pooling has shortened the sequence, use the reduced length; when the channel dimension does not match the LSTM's expectation, apply a 1×1 convolution or linear projection for alignment.
   3. Configure the LSTM with 64 hidden units, tanh activation, and return_sequences=True to retain representations at each time step. Implement this conversion in the model construction module and log the input/output tensor shapes and key settings to ensure reproducibility and traceability, thereby retaining representations at each time step. Implement this conversion in the model construction module and log input/output tensor shapes and key settings to ensure reproducibility and traceability.
   4. Insert a temporal attention block immediately after the LSTM outputs: the block takes the per-time-step hidden representations and uses a single-hidden-layer projection to produce an attention weight vector.
   5. Set the weight-vector dimension to 64, specified directly in the attention layer's configuration. Then, normalize the weights across the time axis with a softmax, yielding the relative importance of each time step. Use the normalized weights to compute a context vector (weighted sum), which is fused with the sequence representation and fed to the regression head.
   6. Train the attention module end-to-end with the backbone; if padding exists, apply a mask before the softmax to ignore invalid steps.
   7. For the ablation experiments only, split the dataset chronologically into training (80%) and test sets (20%), while conducting a separate five-fold cross-validation to evaluate the overall stability and generalization of the proposed model.
   8. Execute training from the project root directory, with all relative paths anchored to that root. Apply early stopping such that if the validation error does not improve for 10 consecutive epochs, training halts and the best weights are saved to ./checkpoints/best_model.h5.
   9. To record the process, a CSV logger writes per-epoch training/validation losses and key metrics to training_log.csv (columns such as epoch, train_loss, val_loss, metrics, timestamp). Update the best-model file whenever validation improves, ensuring reproducibility and supporting subsequent analysis.
      NOTE: CNN automatically extracts spatial features from input data through local connectivity and weight sharing. Shallow convolutional kernels capture subtle local variations. Spatial features manifest as the concentration distribution and patterns of different gases (such as oxygen and carbon monoxide) at the same moment, including local anomalies and concentration gradients. These reflect the spatial correlations of gases during the coal spontaneous combustion process, which CNNs can effectively identify¹⁷. The feature sequence extracted by CNN is fed into an LSTM, which dynamically models the time series through its gating mechanisms. Temporal features such as gradual temperature rise, fluctuations, and sudden changes in gas concentration, and their sequential order reflect the cumulative temperature and gas reactions during coal spontaneous combustion. LSTM effectively captures these patterns, enhancing prediction accuracy and model stability¹⁸. An attention mechanism is introduced at the LSTM output layer to assign weights to features at each time step. This focuses on critical stages of coal spontaneous combustion temperature, strengthens valuable information, suppresses noise, and improves prediction performance¹⁹ (Figure 2).
2. Design of a model structure optimization method based on SSA
   NOTE: The dynamic optimization strategy proposed here integrates model structure design with hyperparameter optimization and is overall divided into the preparation stage, deformation stage, and formal training stage. This method achieves collaborative adjustment of structural parameters and hyperparameters, enabling the model to maintain structural flexibility while improving performance and adaptability to complex operating conditions.
   1. Preparation phase
      1. Before the SSA-based search, define the hyperparameter/structure search space as a 4-tuple: x_i= L_cnn, L_lstm, L_r, batch_size, L_cnn (number of CNN blocks) and L_lstm (number of stacked LSTM layers) are integer-valued and sampled uniformly from {1,2,3,4} and {1,2,3}, respectively.
      2. Because the optimizer proposes real-valued vectors, map non-integer proposals to the nearest integer using Python's round() rule (ties-to-even) and then clip to [1,4] or [1,3]. Draw the learning rate lris is log-uniformly from the interval [1 x 10^-2, 1 x 10^-1]. Clip out-of-range proposals to the nearest bound. batch_size is a discrete choice from {32, 64, 96, 128}.
      3. Unless otherwise stated, do not impose cross-parameter constraints during this preparation step. For reproducibility, apply a common random seed (42) to Python, NumPy, the deep-learning framework, and the environment variable PYTHONHASHSEED.
      4. Initialize the SSA search with a population size of 30 and run for 80 iterations.
         NOTE: These settings, together with the rounding/clipping rules above, defined the preparation of the search space used for all subsequent experiments.
   2. Deformation phase
      1. Randomly generate an initial parameter set, denoted as C₁,L₁,I₁,b₁. Before training the network, round off the integer dimensions to the nearest integer and clip all values to their bounds.
      2. Use a single training/validation run to compute the validation Mean Squared Error (MSE) as the fitness; record and store the current result.
      3. Update the position to generate a new parameter combination C₂,L₂,I₂,b₂. Run a single training iteration and compare its fitness with the stored value.
      4. If the new combination performs better than the previous one, replace the original result and set the current position to C₂,L₂,I₂,b₂. If it performs worse, set a marker to avoid repeatedly selecting ineffective combinations while retaining the incumbent.
      5. Repeat steps 2.2.2.2-2.2.2.4 for multiple iterations until no further improvement is observed, thereby obtaining the optimal combination C_n,L_n,I_n,b_n.
      6. Save C_n,L_n,I_n,b_n as the final structural parameters, and begin formal training.
   3. Training phase
      NOTE: The CNN and LSTM layer counts with the best fitness are selected as the final network configuration for full training and evaluation on the test set (Figure 3).
      1. Set the search range for model structures and hyperparameters, including 2-6 CNN layers, 1-4 LSTM layers, a learning rate range from 1 × 10^-5 to 1 × 10^-2, batch sizes of 32, 64, or 128, and a maximum of 100 training epochs. Optimize these parameters jointly during the SSA iterations.
      2. Initialize the Sparrow Search Algorithm (SSA) population by setting the population size to 30 and the maximum number of iterations to 100. Ensure that each individual represents a candidate model configuration, including CNN depth, LSTM depth, learning rate, and batch size.
      3. In each iteration, divide the SSA population into discoverers (20%), followers (70%), and sentinels (10%). According to the SSA position update rules, discoverers perform global exploration, followers conduct local exploitation, and sentinels prevent the algorithm from being trapped in local optima. Update the position vectors of all individuals after each iteration.
      4. Use the mean squared error (MSE) on the validation set as the fitness function to evaluate the predictive performance of each candidate model. Dynamically adjust the search direction based on the fitness values, allowing SSA to gradually converge toward the optimal structure and hyperparameter configuration.
      5. After completing the SSA iterations, output the optimal CNN depth, LSTM depth, learning rate, and batch size. Retrain the model on the full training set using these optimal parameters and save the final trained model weights to final_model.h5".
         NOTE: The optimization algorithms are widely applied in industrial fields^20,21, commonly used in production scheduling²², quality control²³, equipment maintenance²⁴, resource allocation²⁵, and process parameter optimization²⁶, among others. The sparrow search algorithm (SSA) adopted in this study is an intelligent optimization algorithm simulating the foraging behavior of sparrow populations. It achieves efficient optimization through the collaborative mechanism of discoverers, followers, and sentinels²⁷. The algorithm designates the best individual in the population as the discoverer to perform global exploration, while the remaining individuals act as followers conducting local exploitation, and sentinels are set to avoid local optima²⁸. SSA employs an adaptive strategy to balance exploration and exploitation capabilities, featuring fast convergence speed and simple parameter settings²⁹. In the coal spontaneous combustion temperature prediction model proposed in this paper, SSA serves as the core method of the "structural optimization" concept, automatically optimizing the CNN-LSTM-Attention model architecture and key hyperparameters to enhance prediction accuracy and generalization across heterogeneous datasets.

3. Model validation and transferability evaluation

1. Model effectiveness validation
   1. Design an ablation experiment using coal spontaneous combustion experimental data to verify the individual contributions of the CNN, LSTM, and Attention modules. Divide the dataset randomly into training (80%) and test sets (20%) using a fixed random seed of 42 to ensure reproducibility.
      NOTE: Use a workstation equipped with an NVIDIA RTX 4090 GPU and run all experiments with Python 3.8 and TensorFlow 2.6 (development IDE information is listed in the Table of Materials).
   2. Set the objective function to mean squared error (MSE) and apply the improved Sparrow Search Algorithm (SSA) to jointly optimize key hyperparameters of the CNN-LSTM-Attention architecture, including network depths, learning rate, and batch size. To ensure fairness, train all baseline models on the same dataset using identical training epochs, learning rates, and batch sizes, and evaluate on the same test set.
   3. Based on a fixed CNN-LSTM-Attention model structure, employ five classical parameter optimization methods-Genetic Algorithm (GA), Grey Wolf Optimization (GWO), Particle Swarm Optimization (PSO), Whale Optimization Algorithm (WOA), and Sparrow Search Algorithm (SSA)- to optimize the model, followed by a comprehensive evaluation integrating structural optimization.
   4. Calculate the performance metrics for each model on the test set and visualize the results using comparison charts.
      1. Report the ablation study (Table 1), the predictive performance of the seven models (Table 2), and the comparative performance of optimization algorithms (Table 3).
      2. For the figures, plot model predictions against observations and set the legend labels to "Predicted" and "Measured". Ensure consistent fonts, color schemes, and axis units across all figures. For the tables, keep the typeface, column order, units, and decimal places consistent.
2. Model transferability validation
   1. Constructing the transfer validation dataset
      1. Use a transfer-validation dataset of 83 coal samples from six mining sites, totaling 12,450 temperature-gas feature records. For each site, split the data into training (80%) and test (20%) sets, and maintain balanced representation across diverse geological conditions.
         NOTE: To evaluate the impact of "structural optimization" on the model's generalization performance, this study constructed a comprehensive coal spontaneous combustion temperature feature dataset encompassing multiple mining sites and diverse geological conditions. The dataset integrates field monitoring data from representative mines such as Qinglong Mine (near-horizontal coal seams), Xiao Ji Han Mine (thick coal seams), and Zhangjiamao Mine (shallow buried coal seams), as well as experimental data from I-II class self-ignition prone coal seams under varied geological conditions including the No. 4 coal seam (high volatile content), Yuan Dian No. 2 Mine's 72 coal seam (composite roof), and Hongqingliang Mine (oxidation-prone characteristics).
   2. Cross-mine dataset heterogeneity analysis
      1. Aggregate per-mine records for CO, CO₂, CH₄, C₂H₆, C₂H₄, and coal temperature; keep mine id and timestamps.
      2. Harmonize units (ppm or %) and align timestamps; handle missing values as specified in step 2.1.
      3. Compute min, Q1, median, Q3, max, and IQR for each variable × mine; flag outliers by the 1.5×IQR rule.
      4. Quantify cross-mine heterogeneity by calculating fold differences for Q3 and max; highlight cases ≈ two orders of magnitude (esp. CO/CO₂).
      5. Plot box-and-whisker charts by mine (one panel per variable; common y-axis per variable; show outliers as points; label quartiles).
      6. Export as Figure 4.
      7. Select four well-performing models in the laboratory dataset-XGBoost, BP, TCN, and Transformer-for comparison to validate the transferability of the proposed model.

Results

Model stability was validated using four independent datasets, demonstrating consistent predictive performance across different geological conditions. This section presents representative results and performance evaluations of the coal spontaneous combustion experiments and the proposed SSA-CNN-LSTM-Attention model. First, the variations in multiple gas indicators collected during the programmed heating oxidation experiments are analyzed to reveal the dynamic patterns of gas concentration...

Discussion

Use intact coal blocks; perform only surface cleaning after sampling and double-seal immediately to avoid prolonged exposure. Keep the gas line leak-tight with constant-flow control (MFC), execute the programmed heating exactly as specified in the protocol, and calibrate the GC with certified standards. Acquire temperature and gas signals at fixed intervals and synchronize timestamps (see protocol). From a computational standpoint, fix and record the environment (OS, Python, deep-learning framework, CUDA, etc.), set rand...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
6-port, 2-position gas sampling valve w/ universal actuator	VICI Valco	EUDA-2C6UWT	1/16" fittings; 0.75 mm ports; RS-232; 2" standoff
Alumina sample boat	MTI Corporation	EQ-CA-L50W40H20	>99.5% Al2O3; 50×40×20 mm
Chromatography software	Agilent Technologies	OpenLab CDS	Data acquisition/processing
Coal samples (field-collected)	In-house/field-collected	N/A	Source detailed in Methods
CUDA / cuDNN	NVIDIA	https://developer.nvidia.com/cuda-zone	CUDA 11.x; matching cuDNN
DAQ mainframe	Keysight	DAQ970A	6½-digit DMM; USB/LAN
Desiccant (indicating)	W.A. Hammond Drierite	23001	Calcium sulfate; 8 mesh; 1 lb
Diaphragm air pump	KNF	https://www.knf.com	Continuous air supply; adjustable flow
FR lab coat	Bulwark	KEL2 (series)	NFPA 2112 compliant
Gas chromatograph	Agilent Technologies	G3540A (8890 GC System)	GC system; EPC; up to 2 inlets / 4 detectors
Heat-resistant gloves	Ansell	43-113	Intermittent up to ~350 °C
High-purity air	Air Liquide / Airgas	https://www.airgas.com/solutions/specialty-gases/pure-gases/alphagaz	≥99.99% purity
IDE	JetBrains / Microsoft	https://www.jetbrains.com/pycharm/ ; https://code.visualstudio.com/download	PyCharm / VS Code
Inline filters	Swagelok	https://www.swagelok.com/downloads/webcatalogs/en/ms-01-92.pdf	Sintered SS elements 0.5–15 µm
K-type thermocouples	OMEGA	https://www.omega.com	Type K (NiCr–NiSi)
Mass flow controller (0–200 sccm)	Alicat Scientific	MC-200SCCM-D	MC-series; ±(0.8% rdg + 0.2% FS)
Multi-component calibration gases	Messer	https://specialtygases.messergroup.com/standard-gas-mixtures	Custom concentration; certificate
Operating system	Microsoft	https://www.microsoft.com/en-us/software-download/windows11	Windows 11
Operating system	Canonical	https://ubuntu.com/download/desktop	Ubuntu LTS (22.04/24.04)
PTFE/PFA tubing	Swagelok	https://products.swagelok.com/en/all-products/hoses-flexible-tubing/ptfe-pfa-core-hose/c/716?clp=true	Chemically resistant; 1/16–1/4 in OD
Python	Python Software Foundation	https://www.python.org/downloads/	Version 3.8
Quartz sample boat	MTI Corporation	EQ-QB-1017 (example size)	~1200 °C working temp
Safety goggles	3M	93506P1-DC (example)	Chemical splash; anti-fog options
Stainless-steel seamless tubing	Swagelok	https://www.swagelok.com/downloads/webcatalogs/en/ms-01-181.pdf	316/316L; 1/16–1/4 in OD
Stainless-steel tube fittings & ferrules	Swagelok	https://products.swagelok.com/en/all-products/fittings/tube-fittings-adapters/c/154?clp=true	316/316L; double-ferrule
Temperature controller	Eurotherm	3216	Single-loop PID; programmable ramps/alarms
TensorFlow	Google	https://www.tensorflow.org	Version 2.6
USB thermocouple module	NI (National Instruments)	781314-01 (USB-TC01)	K/J/T; logging software
Variable area flowmeter (Visi-Float)	Dwyer Instruments	VFA-2-EC-SS (0.2–2 SCFH Air)	Low-flow range; direct reading
Workstation GPU	NVIDIA	900-1G136-2530-000 (Founders Edition)	GeForce RTX 4090, 24 GB GDDR6X (FE)