Research on Characteristic Analysis of Typical Fire Accidents and Trend Prediction of Workplace Accidents

Abstract

Workplace accidents pose a serious threat to people’s lives and property and hinder social and economic development. Among these accidents, fire accidents are typical due to their sudden occurrence and severe consequences. To better understand accident evolution laws and improve risk prevention, this study analyzes the characteristics of typical national fire accidents based on 2015–2024 accident statistics. A linear-nonlinear combined Autoregressive Integrated Moving Average-Long Short-Term Memory (ARIMA-LSTM) model is established to predict trends of the number of national overall workplace accidents, deaths, injuries, and direct economic losses, and it is compared with Autoregressive Integrated Moving Average (ARIMA), Long Short-Term Memory (LSTM), Seasonal Autoregressive Integrated Moving Average (SARIMA), and Seasonal Autoregressive Integrated Moving Average-Long Short-Term Memory (SARIMA-LSTM) models. The results show that the ARIMA-LSTM model integrates the strengths of linear fitting and nonlinear learning, with stronger explanatory power and higher prediction accuracy, as reflected by lower Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE) values. This study provides technical support for the precise prevention and control of fire accidents, trend prediction of work safety accidents, and helps to establish a scientific and forward-looking safety risk prevention and control system.

1. Introduction

Work safety, as a core component of the national security system, is directly related to the lives and property safety of the people, social harmony and stability, and high-quality economic development, and serves as an important support for advancing the modernization of the national governance system and governance capacity. Currently, the overall situation of work safety in China is stable with a positive trend and continues to improve, but the total number of accidents remains at a high level. The risk of occasional and frequent major and extraordinarily serious accidents has not been eradicated, and work safety governance still faces practical challenges such as numerous and widespread risk points, complex and diverse causes, and great difficulty in prevention and control.

As a typical and highly hazardous type of accident in the field of work safety, fire accidents are characterized by strong suddenness, fast spread speed, wide coverage, and high disposal difficulty [1]. They are prone to causing mass casualties and heavy property losses and have always been the focus of work safety supervision, risk prevention, and control. From the profound pain of the “8·12” Extraordinarily Serious Fire and Explosion Accident at Tianjin Port to the tragic lesson of the “11·21” Extraordinarily Serious Fire Accident in Anyang, Henan, from the widespread social concern caused by the “4·18” Major Fire Accident at Changfeng Hospital in Fengtai, Beijing, to the Level 5 Major Fire that occurred in Hong Kong on 26 November 2025, each typical accident has profoundly warned of the importance and urgency of fire prevention and control [2].

Therefore, conducting an analysis of the characteristics and laws of fire accidents, while accurately predicting the overall development trend of workplace accidents, is of great theoretical value and practical significance for improving accident prevention capabilities and reducing accident risks [3]. It is urgent to construct scientific and efficient prediction models and technical methods to provide strong support for the transformation of work safety governance to pre-accident prevention [4,5].

In recent years, the analysis and prediction of workplace accidents have become a prominent research focus in occupational safety. Driven by advances in data-driven technologies, accident prediction research has evolved from traditional experience-based approaches toward intelligent modeling, resulting in a rich ecosystem of prediction models tailored to diverse scenarios. According to modeling principles, existing accident prediction methods can be categorized into three major groups: statistical analysis models, machine learning models, and deep learning models. Each category exhibits distinct strengths and applicability based on its underlying techniques. Meanwhile, research trends increasingly emphasize model integration and interpretability, with applications expanding across high-risk industries.

Statistical analysis models represent the traditional foundation of accident prediction. Rooted in mathematical statistics, they focus on capturing linear relationships and perform well for datasets of moderate size with simple feature structures. Their key advantages include transparent mechanisms, high interpretability, and low computational requirements. Recent studies demonstrate the continuous development and refinement of statistical approaches. Sharaf A and Moudhi A [6] developed a Bayesian hierarchical statistical model to forecast traffic accident counts using five years of data from Abu Dhabi, achieving a prediction accuracy of 72% and providing actionable insights for traffic planning. Ordysiński [7] combined univariate and multivariate techniques to propose a novel statistical framework for predicting work-related injury severity using the European Statistics on Accidents at Work (ESAW) database; the model attained over 80% accuracy in binary classification, supporting occupational risk management. Hong et al. [8] integrated regression analysis with roadway geometry and traffic volume data to build urban accident prediction models, enabling accident rate estimation for new or reconstructed road networks. Dhamaniya [9] applied regression methods to construct two-tier accident prediction models under mixed traffic conditions, incorporating a driver–pedestrian index to improve model specificity. Qu et al. [10] enhanced the Grey–Markov model into a multi-factor framework, improving medium-term forecasting performance for highway accidents. Yang et al. [11] adopted a Bayesian negative binomial (B-NB) regression model for expressway tunnel accident prediction, identifying six critical risk factors and achieving 76.1% accuracy. Pandey et al. [12] employed Bayesian networks to analyze steel industry accidents, explicitly uncovering machine-related and human-error contributors to incidents.

Machine learning models, centered on data-driven learning, do not require explicit assumptions about data distributions and excel at capturing nonlinear patterns and generalizing to unseen scenarios. They have become the mainstream approach in accident prediction, encompassing ensemble learning, classical classification/regression, and interpretable variants. Ensemble methods are particularly prevalent due to their high accuracy and robustness. Xia et al. [13] constructed safety accident cost prediction models using random forest, extreme gradient boosting (XGBoost), and natural gradient boosting (NGBoost), based on 1606 construction industry accident records. SHapley Additive exPlanations (SHAP) values were applied to enhance interpretability, and the random forest regression model was identified as optimal. Shin [14] applied the stochastic gradient boosting (SGB) method to early construction site accident prediction; comparisons with artificial neural networks (ANNs) and decision trees (DTs) confirmed SGB’s advantages in parameter simplicity, interpretability, and predictive performance. Chen et al. [15] integrated text mining, XGBoost, and SHAP analysis to develop a bus accident severity prediction framework using over 15,000 records from a city in Jiangsu Province, substantially improving both accuracy and interpretability. Zhang et al. [16] combined XGBoost and SHAP to build a maritime accident classification model using Zhejiang Maritime Bureau data, achieving 97.14% accuracy. Brandt et al. [17] incorporated meteorological features into machine learning models, raising cross-validation accuracy from 64.86% to 70.23% and identifying wind speed, sea-level pressure, and visibility as key predictors. Bu [18] designed a multi-stage fusion strategy integrating multilayer perceptron (MLP) and random forest for traffic accident severity prediction, reaching 94% accuracy. Chen et al. [19] proposed an MSCPO-XGBoost hybrid model with optimized hyperparameters, outperforming baseline algorithms in comprehensive performance. Additionally, Kim et al. [20] developed an interpretable maritime accident prediction system using classical machine learning, validated via the Service Quality (SERVQUAL) framework. Kumar et al. [21] applied a Conditional Generative Adversarial Network-Extreme Boosting (CGAN-EB) approach integrating geometric and traffic parameters, achieving 98% accuracy for four-lane-divided-highway accident prediction.

Deep learning models leverage hierarchical network architectures to capture complex nonlinear interactions and spatiotemporal dependencies, making them suitable for large-scale, high-dimensional datasets. They address critical limitations of traditional models, including small-sample scarcity and class imbalance. Xiong et al. [22] developed a multi-task deep convolutional neural network (MTCNN) integrated with Bayesian optimization for multi-dimensional aviation accident prediction, mitigating class imbalance to improve performance. Zeng et al. [23] enhanced a Wasserstein conditional Generative Adversarial Network (WCGAN) with NeuralProphet and gradient penalty, alleviating overfitting and mode collapse in small-sample aviation accident forecasting and outperforming baselines by over 16%. Bae et al. [24] implemented Transformer and Mamba architectures as surrogate models for real-time prediction of severe nuclear power plant accidents, supporting operator decision-making under instrument failure conditions. Yin et al. [25] combined word embedding and attention mechanisms for interpretable prediction of highway–rail grade crossing accidents. Harada et al. [26] proposed an object-detection-free deep learning model specifically for high-fatality single-vehicle accident prediction, demonstrating strong performance on dedicated datasets. Li et al. [27] developed an Attention Spatial-Temporal Multi-Graph Convolutional Network (ASTMGCN) to capture dynamic traffic spatiotemporal patterns, addressing data sparsity and black-box limitations. Hu et al. [28] integrated large language models (LLMs) with cloud–edge networks to construct a multi-dimensional traffic risk prediction framework, surpassing conventional machine learning approaches. Zhao et al. [29] proposed a knowledge graph-based hazardous chemical accident prediction method (KG-HCAP), reformulating prediction as semi-inductive link prediction and achieving 91.7% accuracy for accident type classification.

Overall, statistical, machine learning, and deep learning models form a complementary and hierarchical framework for work accident analysis and prediction. Intelligent modeling and interpretability enhancement remain dominant research directions. Nevertheless, the current literature exhibits notable limitations. First, fire accidents—a cross-industry, high-frequency work safety hazard—are rarely the primary focus of systematic investigation. Most studies treat fires as secondary events, lacking a comprehensive analysis of statistical patterns, including temporal distribution, severity structure, and industrial variability. Second, for general workplace accident forecasting, traditional statistical models fail to capture non-stationarity and long-term dependence, two defining characteristics of accident time series. Single-model approaches also suffer from common drawbacks: limited linear fitting, difficult parameter optimization, and inadequate nonlinear learning capacity, rendering them unsuitable for long-term accident forecasting.

Based on the above research limitations, this paper focuses on typical fire accidents to carry out statistical characteristic analysis of basic data and considers the linear trends and nonlinear fluctuation characteristics of accident data of different scenarios. A linear-nonlinear decomposition modeling idea is adopted to construct an ARIMA-LSTM combined prediction model for the trend prediction of national overall workplace accidents. The ARIMA model is used to accurately extract linear-trend components, and LSTM is used to learn the nonlinear fluctuation characteristics in residual sequences, realizing complementary advantages, improving the fitting and prediction ability of complex data, and providing more reliable technical support for the prevention of workplace accidents.

2. Materials and Methods

2.1. Research Data

2.1.1. Data Source

This study collected and analyzed investigation reports on typical workplace accidents in China over the past decade. The data collection period was from January 2015 to December 2024, and all official records were verified and finally confirmed in December 2025. The data are obtained from the following sources:

(1)

Official websites of provincial, autonomous regional, and municipal governments and their departments of emergency management;

(2)

Official websites of prefecture-level city governments and local emergency management bureaus;

(3)

The China Work Safety Big Data Platform.

Data screening strictly follows the principles of authenticity, typicality, and completeness. Preference is given to workplace accident cases with complete official investigation reports and full information. Records with incomplete information, unclear casualties, or missing loss data are excluded. Typical workplace accidents, including fire accidents, industrial, mining, and commercial accidents, and construction accidents, are mainly included. Finally, a valid analysis dataset is established to provide a data foundation for the subsequent characteristic analysis and model prediction.

2.1.2. Basic Data Information

A total of 4542 investigation reports of typical workplace accidents from 2015 to 2024 were collected, including 3083 minor accidents, 1294 serious accidents, 148 major accidents, and 17 catastrophic accidents. The annual distribution and accident grade distribution of the accident investigation reports are shown in

Table 1. Distribution of years and accident levels in accident investigation reports.

Year	Number of Minor Accident Reports	Number of Serious Accident Reports	Number of Major Accident Reports	Number of Catastrophic Accident Reports
2015	324	152	27	5
2016	475	151	24	4
2017	587	150	21	1
2018	422	215	16	0
2019	281	204	18	2
2020	262	151	13	0
2021	292	124	13	0
2022	233	79	7	2
2023	144	55	8	2
2024	63	13	1	1
Total	3083	1294	148	17

.

2.2. Research Models

2.2.1. ARIMA Model

The ARIMA model [30] is a classic traditional time-series analysis model. It converts non-stationary sequences into stationary ones through differencing (integration) and captures linear dependencies by combining autoregression (AR) and moving average (MA). Its core ideas include: (1) differencing stationarization—performing $\mathrm{d}$-order differencing on the original sequence ${\mathrm{y}}_{\mathrm{t}}$ to obtain a stationary sequence ${∆}^d{y}_t$, where $∆y_t=y_t-y_{t-1}$; and (2) linear modeling—constructing a combined $\mathrm{A}\mathrm{R}\left(\mathrm{p}\right)$ and $\mathrm{M}\mathrm{A}\left(\mathrm{q}\right)$ model for the stationary sequence to describe the linear relationship between the current value, historical values, and historical errors.

The general form of the ARIMA model is

$$\varnothing \left(B\right){(1-B)}^d{y}_t=\theta {\left(B\right)}_{{ϵ}_t}$$

(1)

where $B$ is the lag operator ($B_{y_t}=y_{t-1}$); $\varnothing \left(B\right)=1-{\varnothing }_{1}B-\cdots -{\varnothing }_p{B}^p$ is the $p$-order autoregressive polynomial; and $\theta \left(B\right)=1-{\theta }_{1}B-\cdots -{\theta }_q{B}^q$ is the $q$-order moving average polynomial; ${ϵ}_t$ is the white noise sequence. When $\mathrm{d}=0$, the model degrades to a standard ARMA model; when $p=0$ or $q=0$, it degrades to an IMA or AR model, respectively.

2.2.2. LSTM Model

LSTM [31] is an improved variant of Recurrent Neural Networks (RNNs). It solves the gradient vanishing and long-term dependency problems of traditional RNNs by introducing a gating mechanism. The structure of the LSTM unit is shown in Figure 1.

The core structure of the LSTM model includes an input gate, a forget gate, and an output gate. The input gate controls the input of new information, the forget gate determines whether to retain or discard old information in the cell state, and the output gate defines the output value. Through this mechanism, LSTM can effectively process long time-series data, avoid gradient vanishing and explosion problems, and better learn long-term dependencies in data.

2.2.3. ARIMA-LSTM Model

The ARIMA-LSTM model [32] combines the advantages of two classic algorithms. ARIMA is adopted to capture linear trends in time-series data, while LSTM effectively mines complex nonlinear fluctuation rules.

(1)

The ARIMA model is used to capture linear temporal characteristics of time series and is suitable for processing stationary or differenced stationary data. In the combined model, ARIMA fits the linear part of the data, leaving nonlinear or complex dependencies to the LSTM model.

Implementation steps:

1)

Training set division: Split the data into training and test sets at a certain ratio.

2)

Stationarity test: Verify the stationarity of the original sequence through the Augmented Dickey–Fuller (ADF) test to determine whether differencing is required.

3)

Parameter setting: Determine the model parameters $(\mathrm{p},\mathrm{d},\mathrm{q})$ based on the Partial Autocorrelation Function (PACF) truncation characteristics and Bayesian Information Criterion (BIC).

4)

Linear prediction and residual extraction: Predict the test set to obtain the linear-trend prediction value ${y^}_{ARIMA}$. The residual sequence is defined as: $e_t=y_t-{y^}_{ARIMA}$. The residuals contain nonlinear features not captured by the linear model and are used as input to the LSTM model.

(2)

The LSTM model excels at learning nonlinear dependencies and long-term memory features of time series. In the combined model, LSTM fits the nonlinear part of ARIMA residuals to compensate for ARIMA’s insufficiency in modeling complex patterns.

Implementation steps:

1)

Data standardization: Normalize residuals to the range [0, 1] using MinMaxScaler to avoid the impact of numerical differences on training.

$$e_t^{norm}={\displaystyle \frac{e_t-min\left(e\right)}{\mathit{max}\left(e\right)-min\left(e\right)}}$$

(2)

2)

Perform a stationarity test on the residual sequence.

3)

Sequence construction: Update the input sequence with the latest residual predictions through a sliding window mechanism and determine the window length L to avoid information loss.

4)

Network structure: Determine parameters such as input layer, hidden layer, output layer, activation function, training batch size, and step size to build the LSTM network structure. The prediction deviation of ARIMA is corrected by learning the nonlinear patterns of residuals.

(3)

Model integration: The prediction formula of the combined model is

$${y^}_{Combined}={y^}_{ARIMA}+{e^}_{LSTM}$$

(3)

where ${e^}_{LSTM}$ is the residual prediction value of the LSTM model.

3. Characteristic Analysis of Typical National Fire Accidents

Workplace accidents are diverse and complex. Fire accidents occur frequently and suddenly, spread rapidly [33,34], and easily cause mass casualties and major property losses. They are typical accidents that require key prevention in work safety. The occurrence laws and disaster characteristics of fire accidents reflect the key risks for work safety and are closely related to the overall work safety situation.

3.1. Temporal Distribution Characteristics

From 2015 to 2024, a total of 181 fire accident data were collected, accounting for 4% of the total data. As shown in Figure 2, from 2015 to 2024, the number of accidents, casualties, and economic losses of typical national fire accidents shows an obvious temporal trend, which generally rises first and then falls.

In terms of accident number, it increased from 14 in 2015 to a peak of 31 in 2018–2019, then decreased gradually after 2020 and dropped to 5 in 2024. This shows that China’s fire prevention and control have achieved remarkable results. In terms of casualties, the death toll reached a peak in 2015 due to the catastrophic Tianjin Port accident and remained relatively stable in other years. The number of injuries showed a similar trend. In terms of economic loss, it also peaked in 2015 and remained stable with a slight decline in later years.

The accident classification standards adopted in this study strictly follow the “Regulations on the Reporting, Investigation and Handling of Workplace accidents” (State Council Decree No. 493).

Table 2. Monthly statistics of accident quantities.

Months	Number of Accidents	Minor Accident	Catastrophic Accident	Serious Accident	Major Accident
1	16	7	1	8	0
2	15	2	0	10	3
3	13	2	0	11	0
4	19	8	0	8	3
5	15	5	1	8	1
6	17	5	0	9	3
7	16	6	0	9	1
8	20	7	1	10	2
9	15	8	0	5	2
10	11	2	0	7	2
11	14	4	1	6	3
12	10	2	0	4	4
Total	181	58	4	95	24

presents the monthly distribution of accidents. The total monthly number of accidents fluctuated between 10 and 20, with the maximum of 20 occurring in August and the minimum of 10 in December. Four catastrophic accidents were scattered across January, May, August, and November. Serious accidents constituted the predominant category in all months. Minor accidents reached their highest count of eight in September. Major accidents were reported in several months, including February, April, and June, while no major accidents occurred in January.

3.2. Accident Grade Distribution Characteristics

As shown in Figure 3 and Figure 4, typical fire accidents show an obvious gradient distribution dominated by serious accidents and minor accidents. Major and catastrophic accidents are fewer but extremely harmful.

Serious accidents account for the highest proportion (52.5%), followed by minor accidents (32.0%), major accidents (13.3%), and catastrophic accidents (2.2%).

Although catastrophic accidents are the fewest, they cause the most serious casualties and economic losses. For example, the “8·12” Tianjin Port fire and explosion accident in 2015 caused 165 deaths, 798 injuries, and direct economic loss of 6.866 billion yuan, accounting for 98.7% of the total loss in 2015.

3.3. Industrial Distribution Characteristics

As shown in Figure 5, fire accidents are highly concentrated in four industries: commerce and trade, light industry, chemical industry, and construction.

Commerce and trade have the highest frequency of accidents, which occur mostly in crowded places such as shopping malls, markets, supermarkets, and hospitals. The light industry mainly involves wood processing, furniture manufacturing, and textiles with large amounts of flammable materials. Chemical industry accidents are serious due to flammable and explosive chemicals. Construction accidents mainly occur at construction sites and ongoing projects.

4. Model Calculation and Comparative Analysis

To predict the trends of workplace accidents, core indicators are selected to fully reflect the accident scale, harm degree, and impact scope. Based on the practical needs of work safety supervision and the availability and continuity of data, this paper chooses the number of accidents, accident grade, number of deaths, number of injuries, and direct economic loss as key prediction indicators. These indicators cover accident frequency, casualties, and economic losses—classic metrics for evaluating work safety. Their complete time-series data and obvious trend characteristics effectively support the fitting and prediction of the ARIMA-LSTM model and comparison models, and the results provide a quantitative basis for safety risk prevention and decision-making.

4.1. Data Preprocessing

In the data preprocessing stage, the collected public accident data were systematically cleaned and standardized. First, duplicate records, samples with missing values, and abnormal outliers were removed to ensure data consistency and validity. Next, the time field formatted as “year-month” was parsed into a standard datetime format and set as the dataset index to facilitate time-series analysis. Based on the cleaned time-indexed data, monthly aggregated statistics were computed for the following core indicators: total number of accidents, minor accidents, serious accidents, major accidents, catastrophic accidents, number of deaths, number of injuries, and direct economic loss. The monthly time-series trends of these indicators are visualized in Figure 6.

4.2. Model Calculation

(1)

Parameter Determination and Calculation

1)

Training set division: In total, 80% of the data is used as the training set, and the remaining 20% as the test set.

2)

Stationarity test: The ADF test (p < 0.05) confirms the original sequence is non-stationary and becomes stationary after first-order differencing (d = 1).

3)

ARIMA parameter setting: The ARIMA model parameters are determined as (5,1,0). Specifically, p = 5 represents the autoregressive order confirmed by the truncation feature of PACF; d = 1 denotes adopting first-order difference processing to realize sequence stationarization; q = 0 means there is no obvious moving average feature in residual sequences, and this parameter combination is finally determined based on the minimum BIC.

$$∆y_t={\varphi }_1∆y_{t-1}+\cdots +{\varphi }_5∆y_{t-5}+{ϵ}_t$$

(4)

where $∆y_t$ is the first-order difference sequence, and ${ϵ}_t$ is white noise.

4)

Linear prediction and residual extraction: Predict the test set to obtain the linear-trend prediction value ${y^}_{ARIMA}$.

5)

Residual data standardization: Normalize residuals to [0, 1] using MinMaxScaler.

6)

Perform a stationarity test on the residual sequence, which is a stationary sequence (p < 0.05). At the same time, ACF and PACF have no obvious tail or truncation features, indicating that the residual has basically eliminated linear information, and the remaining part is a stationary nonlinear component, suitable as input for an LSTM model for nonlinear modeling.

7)

Sequence construction: The LSTM predicts based on the latest 10 residuals to avoid information loss. The window length is L = 10, and it is determined via multi-group window comparative cross-validation. We set five candidate window lengths, including 5, 8, 10, 12, and 10 for repeated tests. The verification results show that when L = 10, the model achieves the minimum prediction error and optimal stability, which can effectively balance the extraction efficiency of time-series dependent features and overall computational cost. The i-th sample is

$$X_i=\left[e_i^{norm},e_{i+1}^{norm},\cdots ,e_{i+L-1}^{norm}\right], y_i=e_{i+L}^{norm}$$

(5)

8)

LSTM network structure: The input layer receives sequences of shape (L,1); the LSTM layer has 50 neurons with the Relu activation function (captures long-term dependencies); the output layer outputs 1 value (predicted residual). The model uses the Adam optimizer, MSE loss function, 50 training epochs, and a batch size of 16.

9)

Combined prediction: A recursive prediction strategy is adopted for workplace accident prediction. A rolling recursive strategy is adopted: each step predicts only one month ahead. The predicted value is then fed back into the input window to update the LSTM sequence. This process iterates until all 12 months are predicted, ensuring temporal continuity and stability.

(2)

Calculation Results

The prediction trend of accident characteristic data by the ARIMA-LSTM model is shown in Figure 7. Model prediction represents the fitting result of the model on the test set; 2025 prediction refers to the predicted value for the year 2025.

4.3. Model Comparative Analysis

This paper compares the ARIMA-LSTM model with ARIMA, SARIMA, LSTM, and SARIMA-LSTM models. MSE, RMSE, and MAE are the three most commonly used indicators. They measure the deviation between predicted and actual values from different dimensions, and their combined use can comprehensively evaluate model performance. The index values of each model are shown in

Table 3. Index values of various models.

Prediction Indicator	Model Evaluation Index	ARIMA	SARIMA	LSTM	SARIMA-LSTM	ARIMA-LSTM
Total Number of Accidents	MSE	102.24	79.55	240.8	146.3864	10.0978
	RMSE	10.11	8.92	15.52	12.099	3.1777
	MAE	9.33	8.26	14.11	11.5354	2.9252
Number of Minor Accidents	MSE	31.55	34.14	299.18	49.0948	3.6594
	RMSE	5.62	5.84	17.3	7.0068	1.9130
	MAE	5.06	5.49	16.49	6.4187	1.6154
Number of Serious Accidents	MSE	15.48	21.8	126.64	13.8255	1.3306
	RMSE	3.93	4.67	11.25	3.7183	1.1535
	MAE	3.75	4.32	10.8	3.4187	0.9813
Number of Major Accidents	MSE	0.4	0.4	0.68	0.1797	0.0850
	RMSE	0.63	0.63	0.82	0.4239	0.2915
	MAE	0.63	0.63	0.74	0.373	0.2103
Number of Catastrophic Accidents	MSE	0.08	0.08	0.08	0.0879	0.0742
	RMSE	0.28	0.28	0.29	0.2965	0.2724
	MAE	0.21	0.19	0.14	0.2478	0.1262
Number of Deaths	MSE	1901.82	1587.54	2613.71	1890.4211	391.5299
	RMSE	43.61	39.84	51.12	43.479	19.7871
	MAE	43.04	39.37	48.77	41.3306	13.6351
Number of Injuries	MSE	5402.8	1026.25	5988.46	2937.1623	132.5097
	RMSE	71.01	32.04	77.39	54.1956	11.5113
	MAE	70.45	30.95	59.73	45.6776	8.7090
Direct Economic Loss	MSE	441,099,926.9	184,177,565.4	75,385,363.82	24,261,980.09	8,642,228.3706
	RMSE	21,002.38	13,571.2	8682.47	4925.6451	2939.7667
	MAE	20,770.84	13,210.06	7872.02	4351.3774	2107.4933

.

SARIMA [35] is an extension of the ARIMA model, which incorporates seasonal autoregressive (SAR), seasonal moving average (SMA), and seasonal differencing (D) to capture periodic and seasonal fluctuations in time series. The model is expressed as SARIMA (p, d, q) (P, D, Q)s, where p, d, and q are non-seasonal orders; P, D, and Q are seasonal orders; and s denotes the seasonal period. In this study, SARIMA is introduced to evaluate its ability to fit the periodic characteristics of accident time series.

SARIMA-LSTM [36] is a hybrid forecasting model combining SARIMA and LSTM. SARIMA first extracts linear trends and seasonal components from the time series, and then LSTM fits nonlinear complex fluctuations in the SARIMA residuals. By integrating the strengths of both models, it captures linear, seasonal, and nonlinear patterns simultaneously. In this study, it is used as a benchmark to verify the prediction performance of the proposed ARIMA-LSTM on workplace accident data.

Risk control is a priority in work safety. Priority should be given to MSE/RMSE to ensure the model has the smallest prediction error for major and above accidents. A small MSE value indicates that the model significantly reduces large errors; a small RMSE value directly reflects a significant improvement in prediction accuracy; a small MAE value means the model’s stability is notably enhanced.

As can be seen from Table 2, the ARIMA-LSTM model, based on linear-nonlinear separation modeling, is significantly superior to single models and other combination strategies in this study. Its advantages are twofold: first, it reduces model complexity—ARIMA handles the linear part, reducing the amount of information the LSTM needs to learn; second, it improves generalization ability—the residual sequence is more stable, reducing the risk of overfitting.

5. Conclusions

(1)

Typical national fire accidents from 2015 to 2024 showed an improving trend. They were highly concentrated in four industries: commerce and trade, light industry, chemical industry, and construction. Among them, commerce and trade had the highest accident frequency. Although the number of accidents in the chemical industry was small, the consequences were more serious.

(2)

Workplace accident data have both linear evolution and local nonlinear fluctuation characteristics. The ARIMA-LSTM combined model, through the mechanism of “linear decomposition + nonlinear correction”, adapts to these dual features. Its prediction accuracy is better than that of single models, providing a feasible technical approach for such data prediction.

(3)

Under the accident dataset and verification scenario from 2015 to 2024, the MAE, RMSE, and MSE quantitative error indicators of the ARIMA-LSTM model were all better than those of the comparison models. The model had better adaptability to trend continuity and fluctuation response of accident data and can provide a quantitative reference for analyzing accident evolution laws.

(4)

These predictions enable early identification of high-risk periods and industries for targeted prevention, quantify accident trends to optimize safety resource allocation, and provide data support for policy-making and proactive prevention.

In future research, we will identify and quantify key causal factors for the high fire incidence in commerce and trade and severe consequences in the chemical industry, including personnel density, flammable materials, process risks, and safety management. These factors, together with macro policies and environmental data, will be integrated into the ARIMA-LSTM model to improve prediction robustness. Interpretability analysis will be added to balance model performance and result explanation, providing targeted support for work safety risk prevention.