Temperature prediction methods, devices, and storage media for high-temperature underground cavern complexes
By decomposing and weighting the temperature sequence of high-temperature underground cavern groups, and combining prior geological information and thermal conductivity characteristics, the problem of low accuracy in traditional temperature prediction technology has been solved, achieving more accurate temperature prediction and safe control of high-temperature construction.
Patent Information
- Application Number
- CN202611211446.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-11
- Publication Date
- 2026-11-17
AI Technical Summary
Traditional underground cavern temperature prediction techniques suffer from low accuracy, especially in deep locations where direct measurement is not possible. Existing technologies fail to effectively distinguish the effects of long-term trends, periodic changes, and short-term fluctuations in temperature signals.
By acquiring historical temperature sequences of high-temperature underground caverns and preprocessing them with prior geological information, the temperature sequences are decomposed into long-term, periodic, and short-term components using the empirical mode decomposition method. Coupling weights are calculated based on spatial location relationships and thermal conductivity characteristics, and each component is weighted and corrected. Feature extraction and weighted fusion are performed in a parallel temporal feature extraction channel, and finally, the predicted temperature is verified through geothermal compliance rules.
It improves the accuracy of deep temperature prediction, reduces the interference of shallow disturbances on deep temperature prediction, enhances the accuracy of prediction results and engineering safety, and can automatically correct outliers and trigger high temperature warnings.
Smart Images

Figure CN123115642A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of temperature prediction technology, specifically to a method, device, and storage medium for predicting the temperature of a high-temperature underground cavern complex. Background Technology
[0002] As underground engineering projects extend into deeper strata, high geothermal environments have become a key factor affecting construction safety, structural stability, and operational durability. The internal temperature of underground caverns in high geothermal environments varies significantly with depth and is influenced by a combination of factors, including geothermal gradients, geological structure, construction disturbances, and ventilation conditions. To ensure the safety of workers, the reliability of equipment, and the long-term stability of the caverns, engineering projects typically require temperature prediction at different depths.
[0003] While sensors can be deployed for real-time measurements at shallower or medium depths, deeper locations are often limited by geological conditions, construction difficulties, or excessive costs, making direct measurement impossible. Therefore, temperature prediction is needed for depths that cannot be practically measured.
[0004] Traditional techniques typically model temperatures based on historical temperature sequences at various depths, such as directly training time-series prediction models using multi-depth temperature sequences to infer temperatures at a target depth. However, existing techniques often only utilize the overall variation characteristics of the temperature sequence itself, failing to consider the impact of long-term trends, periodic variations, and short-term fluctuations in the temperature signal on the correlation between temperatures at different depths. This results in limited prediction accuracy when inferring across depths. Therefore, traditional temperature prediction techniques for underground cavern complexes suffer from low accuracy. Summary of the Invention
[0005] The purpose of this application is to provide a method, device, and storage medium for predicting the temperature of underground cavern groups with high geothermal temperatures, in order to solve the problem of low accuracy in traditional temperature prediction technologies for underground cavern groups.
[0006] To achieve the above objectives, the first aspect of this application provides a method for predicting the temperature of high-temperature underground cavern groups, comprising: Historical temperature sequences at different monitoring depths of high-temperature underground caverns are obtained, and the historical temperature sequences are preprocessed in combination with prior geological information, including rock thermal conductivity and geothermal gradient. The preprocessed historical temperature sequence is decomposed into long-term components, periodic components, and short-term components using the empirical mode decomposition method. Based on the spatial relationship and thermal conduction characteristics between different monitoring depths, the coupling weight of each component in the depth direction is calculated, and the corresponding components at each depth are weighted and corrected using the coupling weight to obtain the corrected long-term component, periodic component and short-term component. The corrected long-term component, periodic component, and short-term component are respectively input into three parallel time-series feature extraction channels for feature extraction. The extracted features are then weighted and fused based on the fluctuation contribution of each component to obtain fused time-series features. The fused temporal features are input into the prediction model, the initial predicted temperature of the depth to be measured is output, and the initial predicted temperature is verified by the geothermal compliance rules to output the final target predicted temperature.
[0007] A second aspect of this application provides a temperature prediction device for a high-temperature underground cavern complex, comprising: The acquisition module is used to acquire historical temperature sequences at different monitoring depths of high-temperature underground caverns, and to preprocess the historical temperature sequences in combination with prior geological information, including rock thermal conductivity and geothermal gradient. The decomposition module is used to decompose the preprocessed historical temperature sequence into long-term components, periodic components, and short-term components using the empirical mode decomposition method. The correction module is used to calculate the coupling weight of each component in the depth direction based on the spatial position relationship and heat conduction characteristics between different monitoring depths, and to use the coupling weight to perform weighted correction on the corresponding components at each depth to obtain the corrected long-term component, periodic component and short-term component. The extraction module is used to input the corrected long-term component, the periodic component and the short-term component into three parallel time-series feature extraction channels for feature extraction, and to perform weighted fusion of the extracted features based on the fluctuation contribution of each component to obtain fused time-series features. The output module is used to input the fused temporal features into the prediction model, output the initial predicted temperature of the depth to be measured, verify the initial predicted temperature through geothermal compliance rules, and output the final target predicted temperature.
[0008] A third aspect of this application provides a computer-readable storage medium storing a program that can be loaded by a processor and executed using the above-described method for predicting the temperature of high-temperature underground cavern groups.
[0009] The beneficial effects of this application are: This application first introduces prior geological information on rock thermal conductivity and geothermal gradient to preprocess historical temperature sequences, filtering out anthropogenic thermal disturbances such as ventilation and blasting, compensating for the lag effect of rock strata heat conduction, and purifying the true geothermal time series of the strata. Then, through Empirical Mode Decomposition (EMD), the preprocessed time series is decomposed into long-term components corresponding to deep geothermal evolution, periodic components of periodic working conditions in the tunnel, and short-term components of instantaneous thermal disturbances, accurately separating the multi-scale temperature characteristics of high geothermal strata, avoiding the interference of various heat source signals, and accurately distinguishing the respective contributions of strata geothermal changes and anthropogenic disturbances.
[0010] This application combines the thermal conduction characteristics of the formation to characterize the two types of correlations between steady-state geothermal activity and transient disturbances at deep and shallow measuring points from two dimensions: time-series differential synchronization and component synchronization. It generates layered coupling weights by using a distance attenuation factor that conforms to the heat transfer law of the formation, and completes component weighting correction with the geothermal benchmark at the deepest measurable depth. This restores the true heat conduction logic from deep to shallow, solves the problem of excessive interference of shallow disturbances with deep geothermal trends and imbalance of cross-depth correlations, and effectively reduces the temperature prediction deviation of deep measuring points.
[0011] This application sets up three parallel channels to independently extract the time-series features of three types of components. It relies on the adaptive weighted fusion feature obtained by integrating the proportion of the stratum basement and the fluctuation contribution obtained by instantaneous fluctuations. Then, it combines the depth to be measured into the prediction model. Finally, it conducts geothermal compliance verification through dual thresholds of high geothermal construction temperature and heating rate. When anomalies occur, it automatically corrects and triggers high temperature warnings. The network can adaptively focus on the long-term geothermal core features of the stratum, taking into account both prediction accuracy and the safety management capability of high-temperature construction in deep caverns.
[0012] Other features and advantages of this application will be described in detail in the following detailed description section. Attached Figure Description
[0013] Figure 1 This is a flowchart illustrating a method for predicting the temperature of a high-temperature underground cavern group provided in an embodiment of this application. Figure 2 This is a schematic diagram of the structure of a temperature prediction neural network provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of multiple temporal feature extraction channels provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of a temperature prediction device for a high-temperature underground cavern group provided in an embodiment of this application. Detailed Implementation
[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0015] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified. Details are set forth in the following description for illustrative purposes. It should be understood that those skilled in the art will recognize that this application can be implemented without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid unnecessarily obscuring the description of this application. Therefore, this application is not intended to be limited to the embodiments shown, but rather to be consistent with the broadest scope of the principles and features disclosed herein.
[0016] Figure 1 This is a flowchart illustrating a method for predicting the temperature of a high-temperature underground cavern complex, as provided in an embodiment of this application. Figure 1 As shown, this temperature prediction method may include steps such as 101-105, which will be described in detail below.
[0017] Step 101: Obtain historical temperature sequences at different monitoring depths in high-temperature underground caverns, and preprocess the historical temperature sequences in conjunction with prior geological information.
[0018] In this embodiment, temperature sensors can be uniformly deployed along the vertical burial depth or longitudinal excavation direction of the underground cavern, covering the shallow ventilation zone, the intermediate transitional rock strata, and the deep high-geothermal core area. Long-term temperature time-series data are continuously collected to form independent historical temperature sequences at multiple monitoring depths. Since heat conduction in the rock strata has a time delay, direct modeling can confuse inherent geothermal activity with temporary disturbances, leading to distorted deep temperature prediction results. Therefore, unlike modeling using raw collected data directly, the raw temperature in this step simultaneously includes multiple interference signals from natural geothermal activity, ventilation, and mechanical tunneling blasting.
[0019] Prior geological information refers to the inherent fixed parameters of the strata obtained from geological surveys before cavern excavation. These parameters may include the thermal conductivity of the rock mass and the geothermal gradient. The thermal conductivity of the rock mass characterizes the inherent ability of the rock strata to conduct heat, while the geothermal gradient represents the natural rate of temperature rise of the strata with increasing burial depth. Only deep underground engineering projects with high geothermal temperatures have complete survey data for both.
[0020] As an example, standard theoretical temperature curves for each depth can be constructed based on the global stratigraphic geothermal gradient. The difference between the real-time acquired temperature and the theoretical baseline can be calculated to separate the local thermal deviation sequence caused by ventilation and blasting, filtering out long-term offsets caused by the stratigraphic background temperature baseline. Then, the time lag value of heat passing through the rock strata can be calculated based on the thermal conductivity of the corresponding rock mass, and the deviation time series can be shifted and compensated to eliminate the temporal misalignment caused by heat transfer in the rock strata. The effective time series that only reflects the natural geothermal evolution of the strata can be extracted, reducing the negative impact of shallow temporary disturbances on deep temperature modeling from the data source, and providing clean input data for subsequent multi-scale component decomposition.
[0021] Step 102: Decompose the preprocessed historical temperature sequence into long-term components, periodic components, and short-term components using the empirical mode decomposition method.
[0022] EMD is an adaptive non-stationary time series decomposition algorithm that does not require pre-setting decomposition basis functions. It completes the multi-scale oscillation signal decomposition entirely based on the extreme value characteristics of the time series itself, adapting to the nonlinear and non-stationary geothermal temperature variation laws of underground caverns. For a single pre-processed geothermal time series, all local maxima and minima are identified. Cubic spline interpolation is used to generate upper and lower envelopes, the envelope mean is calculated and removed from the original time series, and this process is repeated iteratively until the screening termination condition is met, yielding multi-level intrinsic mode functions (IMFs) and residual trend terms.
[0023] Then, based on the physical meaning of temperature changes, three components are classified. The first layer of IMF, as the short-term component, represents the rapid temperature fluctuations caused by instantaneous tunneling and ventilation. The residual terms after all iterations are completed are the long-term component, corresponding to the slow natural geothermal evolution of the strata. The values of all remaining IMFs except the first layer are superimposed at each moment to form a periodic component, corresponding to the periodic temperature changes caused by daily ventilation and cyclical construction.
[0024] Temperature variations in underground caverns are influenced by three completely independent heat sources, with vastly different cycles and scales of impact. A uniform temporal model would cause these different variation characteristics to mask each other. Separating these components by physical source allows for complete decoupling of geothermal signals at different scales. Independently distinguishing between long-term geothermal trends, periodic environmental disturbances, and instantaneous construction-induced temperature rises avoids interference between different characteristics and enables the extraction of independent evolutionary patterns for each type of temperature change, thus improving the ability of subsequent modeling to capture deep geothermal trends.
[0025] Step 103: Based on the spatial positional relationship and thermal conduction characteristics between different monitoring depths, calculate the coupling weight of each component in the depth direction, and use the coupling weight to perform weighted correction on the corresponding components at each depth to obtain the corrected long-term component, periodic component and short-term component.
[0026] Geothermal heat is conducted unidirectionally from deep to shallow layers, with the deepest monitoring point exhibiting the most stable geothermal signal. Traditional multi-monitoring point fusion methods rely solely on the distance between monitoring points, failing to consider the geothermal conduction patterns and temporal synchronicity. This can lead to excessively high weights for monitoring points experiencing temporary disturbances in the shallow layers, distorting deep geothermal trends. The coupling weights are used to quantify the temperature time series at any monitoring depth, determining the true contribution of the weights to geothermal changes at the predicted depth. Furthermore, considering both temporal synchronization correlation and spatial heat transfer attenuation within the formation, each component generates its own independent set of weights.
[0027] Hilbert transforms were performed on the original components and the time differences between adjacent components to extract the instantaneous phase. The phase difference between each measuring point and the deepest measurable point was compared to characterize the component synchronization degree (steady-state geothermal matching degree) and the differential synchronization degree (instantaneous thermal disturbance propagation consistency). The average of the two synchronization degrees was taken as the basic correlation value. Then, an exponential distance decay function was constructed based on the difference between the monitoring depth and the depth to be measured, with the influence of geothermal transfer weakening as the distance increases. The basic correlation value was multiplied by the distance decay value and normalized to obtain the coupling weight for each depth. Then, the deepest measurable point was used as the geothermal baseline, and all other depth time series were multiplied by their corresponding weights and superimposed onto the deepest point component to complete the weighted correction.
[0028] Synchronization can characterize two types of depth correlations: long-term geothermal activity and transient disturbances. Distance attenuation closely matches the objective law of heat transfer attenuation in the formation. In this way, the contribution ratio of stable deep geothermal monitoring points can be automatically amplified, the interference weight of shallow temporary disturbance monitoring points can be weakened, the true heat transfer relationship from deep to shallow formations can be restored, the problem of imbalance in cross-depth modeling correlation can be solved, and the temperature prediction error of deep areas without monitoring points can be significantly reduced.
[0029] Step 104: Input the corrected long-term component, periodic component and short-term component into three parallel time-series feature extraction channels for feature extraction, and perform weighted fusion of the extracted features based on the fluctuation contribution of each component to obtain fused time-series features.
[0030] The three parallel temporal feature extraction channels are three independent feature extraction branches with completely identical structures. Each branch only receives a single-scale correction component, and independently mines temporal local fluctuations and long-term dependencies at that scale. The three parallel channels receive long-term, periodic, and short-term correction components respectively, and independently complete local and temporal feature extraction.
[0031] The fluctuation contribution is a normalized weighting coefficient derived from the long-term temperature baseline proportion and instantaneous fluctuation amplitude of the comprehensive components. It is used to adaptively adjust the importance of features at different scales in the fusion process. Instantaneous fluctuation values are obtained by calculating the absolute values of the temperature differences between adjacent time points for the three types of components. These values are then combined with the overall numerical proportion of the components to generate a fluctuation contribution sequence. An element-wise multiplication method is used to scale the output features of each channel using the corresponding contribution, and the resulting summations generate the fused temporal features.
[0032] The three types of components exhibit significant differences in scale and engineering value, with long-term geothermal trends being the core of the prediction, while short-term disturbances are merely local noise. Single-channel networks mix features from all scales, easily masking minute geothermal fluctuations with long-term trends, and fixed fusion weights cannot adapt to the varying strengths of geothermal fluctuations across different formations. Setting up independent channels avoids feature overlap, and dynamic contribution values automatically highlight formation-related geothermal features. This approach preserves multi-scale geothermal temporal details, automatically focuses on long-term geothermal evolution characteristics during model training, weakens transient construction disturbances with no engineering significance, and improves the stability of long-term deep temperature predictions.
[0033] Step 105: Input the fused temporal features into the prediction model, output the initial predicted temperature of the depth to be measured, verify the initial predicted temperature through geothermal compliance rules, and output the final target predicted temperature.
[0034] Geothermal compliance rules are based on the safety regulations for construction of underground caverns in high-temperature areas, establishing a dual judgment standard. This standard includes two constraints: the upper limit of the natural absolute temperature of the stratum and the maximum allowable rate of temperature increase. These constraints are used to determine whether the predicted temperature conforms to the objective laws of geothermal activity and the requirements for high-temperature construction control. If the prediction model only outputs numerical values without geothermal safety verification logic, it is prone to producing false high-temperature predictions that violate the laws of geothermal activity, thus failing to support on-site high-temperature risk management.
[0035] The fused temporal features are input into the prediction model to extract global temporal features. Normalized parameters for the burial depth are simultaneously input, and the data are concatenated to obtain the initial predicted temperature. This initial predicted temperature is the temperature before geothermal compliance verification. The geothermal compliance criteria are retrieved, and the initial temperature value and continuous heating rate are compared. If both indicators do not exceed the standard thresholds, the initial temperature is directly output as the final target temperature. If the temperature or heating rate exceeds the standard, it is determined to be a false high temperature caused by model overfitting. The value is automatically corrected to the safe upper limit, and a high-temperature warning signal is triggered simultaneously, outputting the compliance prediction result.
[0036] Purely data-driven models are prone to fitting short-term ventilation and blasting-induced severe temperature rises, resulting in abnormally high temperatures that do not conform to the inherent geothermal activity of the formation, making them unsuitable for direct use in early warning of high-temperature disasters in caverns. Adding geothermal compliance constraints ensures that the prediction results align with the physical limits of the formation, balancing numerical accuracy with engineering safety requirements.
[0037] In step 101, a theoretical temperature baseline curve for each monitoring depth is first constructed based on the geothermal gradient from prior geological information. This theoretical temperature baseline curve is a continuous curve plotted with each monitoring depth on the horizontal axis and the steady-state stratum temperature without human disturbance on the vertical axis. It is calculated solely from the geothermal gradient and the monitoring depth, representing the standard geothermal temperature at that depth when the stratum is unaffected by ventilation or tunneling, serving as a reference benchmark to distinguish between human-induced thermal disturbances. Based on the geothermal gradient values obtained from the survey, combined with the burial depth values corresponding to each temperature sensor, the standard stratum temperature under undisturbed conditions is calculated point-by-point. A continuous theoretical temperature baseline curve is then generated by fitting the monitoring depth as the horizontal axis and the standard geothermal temperature as the vertical axis, with each burial depth location within the curve corresponding to a unique steady-state baseline temperature.
[0038] For example, to obtain temperature data at different depths, multiple temperature monitoring points are deployed at equal intervals (e.g., 10m) along the vertical burial depth direction (or the longitudinal depth direction of the cavern), and each monitoring point is equipped with a temperature sensor to collect temperature data. Thus, the sensor deployment locations correspond to the "monitoring depths" in this embodiment, and the data at each depth originates from the corresponding temperature monitoring point. The deepest measurable depth refers to the deepest monitoring location within the range of the temperature sensor deployment, i.e., the maximum achievable depth point. For example, if 10 monitoring points are arranged along the vertical burial depth direction, the location corresponding to the last monitoring point is the deepest measurable depth.
[0039] Then, the deviation values between the temperature values at each moment in the historical temperature sequence and the theoretical temperature reference curve are calculated. These deviation values are used as the effective temperature sequence characterizing local thermal anomaly disturbances. The deviation value is the difference sequence between the measured temperature at the same moment and the theoretical reference temperature at the corresponding depth. This difference completely eliminates the long-term background geothermal baseline of the formation, leaving only the local instantaneous and periodic thermal anomaly disturbance signals caused by ventilation, blasting, and equipment operation, which is named the effective temperature sequence. The measured temperature at each moment and each monitoring depth in the original historical temperature sequence is collected. The difference between the measured temperature at the same depth and moment and the standard temperature on the reference curve at that depth is calculated to obtain the temperature deviation value at that moment. The deviation values of all moments are arranged continuously to generate an effective temperature sequence containing only local thermal anomaly disturbances.
[0040] Based on the thermal conductivity of the rock mass in the geological prior information, the thermal conduction time constant of the effective temperature sequence at different rock layer depths is determined. The effective temperature sequence is then subjected to hysteresis compensation using the thermal conduction time constant to obtain the preprocessed historical temperature sequence after eliminating spatial thermal conduction delay.
[0041] The thermal conduction time constant is the lag time required for heat to pass through a rock layer of a specified thickness. It is calculated by combining the rock layer thickness and the thermal conductivity of the rock mass, and is used to correct for time-series shifts in temperature signals caused by heat transfer through the rock layers. The cavern is divided into different rock layer zones, and the thermal conductivity and rock layer thickness parameters of each zone are extracted. The corresponding thermal conduction time constants for these zones are then calculated. Since heat generated must be conducted through the rock layers before being detected by sensors, a time lag phenomenon occurs in the perturbation signal. Different rock layers have different thermal conductivity, resulting in significant differences in the lag time.
[0042] The hysteresis compensation process involves performing a time-shift correction on the entire effective temperature sequence based on the calculated heat conduction time constant. This compensates for the signal time misalignment caused by heat transfer across rock strata, aligning disturbance signals at different depths in the time dimension. The effective temperature sequence is corrected by time shifting according to the time constant of the corresponding rock strata: rock strata with high thermal conductivity and low time constants are shifted slightly, while those with low thermal conductivity and high time constants undergo longer time-shift compensation to eliminate time misalignment caused by heat conduction across rock strata. The resulting compensated historical temperature sequence is then preprocessed.
[0043] This step involves two levels of geological correction. First, the inherent geothermal baseline of the strata is stripped away using the geothermal gradient, and anthropogenic disturbance signals are extracted separately. Then, the heat transfer time series offset is corrected based on the thermal conductivity of the rock mass, eliminating the two types of inherent strata interferences at the data source.
[0044] In step 102, based on the local extreme points of the historical temperature sequence, upper and lower envelopes are constructed by cubic spline interpolation and local mean is calculated. Local mean is eliminated through an iterative sieving process until the preset intrinsic mode function stopping criterion is met, resulting in multiple intrinsic mode functions and residual components arranged in descending order of frequency.
[0045] Local extrema are all local maxima and minima in a single preprocessed historical temperature time series curve. They are the sole basis for EMD adaptive decomposition, requiring no manual pre-setting of decomposition frequency bands and relying entirely on the inherent fluctuation characteristics of the data to divide the signal scale. Cubic spline interpolation is a high-precision interpolation algorithm that uses discrete extrema to fit a continuous smooth curve. Compared to linear interpolation, it can accurately fit the nonlinear fluctuation characteristics of temperature time series and is used to generate upper and lower envelope contours. A single preprocessed historical temperature time series is taken, and all time nodes are traversed to select two sets of discrete data: all local maxima and local minima. Cubic spline interpolation is then applied to the sets of maxima and minima respectively to generate continuous and smooth upper and lower envelope curves, resulting in upper and lower envelope lines that completely encompass the entire fluctuation range of the original temperature time series, representing the upper and lower boundaries of the time series fluctuations.
[0046] The local mean is the arithmetic mean of the upper and lower envelope values at the same time, representing the baseline of the temperature time series average at that location, containing slowly changing low-frequency geothermal trends. The local mean is calculated by taking the upper and lower envelope values time-by-time, and then subtracting the corresponding local mean from the original temperature time series to obtain a sieved oscillation sequence. The entire sieving operation of "extremum extraction - envelope generation by interpolation - mean calculation - subtraction" is repeated for the new oscillation sequence, continuously iterating. After each iteration, it is determined whether the current oscillation sequence meets the IMF's preset stopping criteria. If not, the sieving continues. If it does meet the criteria, an independent IMF component is output.
[0047] The IMF (Integrated Fluid Metric) sieving iterations output single-component time series after meeting the stopping criteria. Each IMF contains only a single oscillation frequency; high-frequency IMFs correspond to instantaneous and drastic temperature fluctuations, while low-frequency IMFs correspond to slow and periodic temperature changes. The IMF stopping criterion is a pre-set convergence judgment standard used to terminate the sieving cycle of a single component, avoiding infinite iteration and ensuring decomposition efficiency and component stability. The error threshold of two consecutive sieving results is commonly used as the judgment basis. The extracted IMFs are stripped from the original time series, and the complete sieving process is performed again on the remaining baseline time series, cyclically extracting multiple IMFs. This process continues until the remaining baseline time series shows no obvious oscillations and no valid extreme points can be identified, at which point all sieving is terminated. The remaining baseline at this point is the residual component. The residual component is the remaining time series after all IMFs have been sieved and stripped, without oscillations or fluctuations, retaining only a monotonous and slow changing trend, corresponding to the long-term steady-state geothermal evolution of the formation. All output IMFs are automatically sorted from high to low oscillation frequency, with the first one being the highest frequency and the last IMF being a low-frequency periodic fluctuation.
[0048] Next, the physical orientation of the three components is performed. The first intrinsic mode function with the highest frequency is extracted as the short-term component characterizing transient thermal disturbances. The short-term component oscillates rapidly with abrupt changes in amplitude, matching only transient thermal disturbances such as blasting and temporary ventilation. The residual component is extracted as the long-term component characterizing the slow evolution trend of the deep original geothermal field. The long-term component only reflects the long-term slow changes in natural geothermal temperature in deep strata undisturbed by human intervention. The remaining intrinsic mode functions (excluding the first intrinsic mode function) are numerically superimposed at the same time point to reconstruct the periodic component characterizing the periodic impact of the cavern environment and construction activities. The periodic component uniformly characterizes the periodic temperature changes caused by daily ventilation and cyclical construction. By orienting the three independent components according to the physical meaning of the heat source, the three types of signals—deep steady-state geothermal trend, periodic construction ventilation fluctuations, and transient tunneling thermal disturbances—are completely decoupled, avoiding the aliasing of multi-scale features that obscure the long-term geothermal evolution of the strata and eliminating mutual interference of features at the signal level.
[0049] The core characteristic of transient thermal disturbances caused by blasting and ventilation in high-temperature underground caverns is abrupt temperature changes. The original time series can mask the synchronous changes in the diffusion of disturbances, and the cross-depth coupling relationship of transient thermal disturbances cannot be quantified solely by the original time series. In step 103, for each component at each monitoring depth, the temperature values at adjacent times are subtracted to obtain a time series difference sequence. The time series difference sequence is a new sequence obtained by subtracting the temperature at adjacent time nodes within a single long-term / periodic / short-term component time series. It only describes the instantaneous rate of temperature change over time and is used to capture the propagation dynamics of transient thermal disturbances such as blasting and ventilation. For any type of long-term / periodic / short-term component and any monitoring depth, all timestamps are traversed, and the temperature at the current time is subtracted from the temperature at the previous time to generate the corresponding time series difference sequence.
[0050] Transient thermal disturbances are the main source of interference in high-temperature caverns. Relying solely on the original phase is insufficient to distinguish between the two entirely different coupling relationships: steady-state geothermal synchronization and transient disturbance synchronization. Furthermore, the degree of disturbance propagation matching cannot be quantified independently. Therefore, a Hilbert transform is performed on the time-series differential sequence to extract the differential instantaneous phase. At the same instant, the differential instantaneous phase difference between other monitoring depths and the deepest measurable depth is calculated to obtain the differential synchronization degree, which characterizes the consistency of the transient thermal disturbance propagation trend.
[0051] The Hilbert transform is a time-series complex analytical transform algorithm that converts real-valued temperature signals into complex signals. It extracts phase information from these complex signals to quantify the degree of synchronization deviation in time-series fluctuations, serving as a standard tool for measuring the matching degree between two temperature time series. The differential instantaneous phase is the phase value extracted after performing a Hilbert transform on the time-series differential sequence, reflecting the temporal characteristics of the dynamic change rhythm of thermal disturbances. The Hilbert complex analytical transform is performed on the time-series differential sequence generated in the previous step to extract the differential instantaneous phase corresponding to each moment. Using the phase of the sensor at the deepest measurable depth as a benchmark, the differential phases at all other monitoring depths at the same moment are subtracted and normalized to obtain the differential synchronization degree. The differential synchronization degree is a normalized index of the difference between the differential instantaneous phase of the target monitoring depth and the deepest measurable depth at the same moment. A value closer to 1 indicates a high degree of consistency in the occurrence and diffusion rhythm of transient thermal disturbances at two locations, specifically characterizing the consistency of the propagation trend of transient thermal disturbances. Establish correlation quantitative indicators for transient thermal disturbance dimensions to accurately reflect whether the occurrence and diffusion rhythm of sudden thermal disturbances such as blasting and ventilation at various depths are synchronized, and independently characterize the strength of cross-stratum propagation coupling of dynamic thermal disturbances.
[0052] As an example, differential synchronization satisfies the following formula: ; in, For the first At the depth and at the deepest measurable point The time-difference synchronization degree, For the first At the depth and at the deepest measurable point The difference in instantaneous phase at time, It is a natural constant. For the length of time, The imaginary unit, This is a modulo operation.
[0053] A single phase index cannot distinguish static steady-state geothermal correlations. Deep strata natural geothermal activity is characterized by slow, steady-state changes, not by differential signals. Differential phase analysis alone can only capture abrupt disturbances and cannot characterize the long-term, slow geothermal equilibrium matching patterns of the strata, thus losing the core geothermal coupling characteristics of high-temperature caverns. Therefore, this step can also directly perform Hilbert transform on each component at each monitoring depth to extract the component's instantaneous phase. The component's instantaneous phase is the phase value extracted directly from the original component time series by performing a Hilbert transform, without differential processing, characterizing the overall phase matching relationship of long-term steady-state temperature fluctuations. At the same moment, the component's instantaneous phase difference between other monitoring depths and the deepest measurable depth is calculated. A Hilbert transform is directly performed on the original long-term / periodic / short-term component time series to extract the component's instantaneous phase. Using the phase of the deepest measurable depth as a benchmark, the phase differences at other depths at the same moment are calculated and normalized to obtain the component synchronization degree characterizing the steady-state thermal equilibrium correlation. Component synchronization is a normalized index of the instantaneous phase difference between the target monitoring depth and the deepest measurable depth at the same moment. The closer the value is to 1, the more similar the long-term geothermal steady-state change patterns of the two strata are, characterizing the strength of the correlation in the steady-state thermal balance of the strata. A new steady-state geothermal dimension correlation index has been added to separately characterize the long-term temperature evolution and the matching degree of the steady-state geothermal field of strata at different depths, fully capturing the inherent coupling relationship of deep geothermal energy that is not affected by instantaneous disturbances.
[0054] As an example, component synchronization satisfies the following formula: ; in, For the first At the depth and at the deepest measurable point The degree of synchronization of components at any given moment. For the first At the depth and at the deepest measurable point The difference in instantaneous phase of the components at time t. It is a natural constant. For the length of time, The imaginary unit, This is a modulo operation.
[0055] Next, the differential synchronization degree and component synchronization degree are averaged to obtain the basic synchronization degree. The basic synchronization degree is the arithmetic mean of the differential synchronization degree and component synchronization degree at the same depth and time. It also incorporates two types of time-series correlation characteristics: transient disturbance and steady-state geothermal activity, to comprehensively reflect the degree of geothermal coupling between measuring points. Within the same period, the differential synchronization degree and component synchronization degree at the monitoring depth are averaged to obtain the synchronization value at each depth. That is, the differential synchronization degree and component synchronization degree at the same depth and time are added together and divided by 2 to obtain the basic synchronization degree. Synchronization incorporates two types of cross-depth correlation characteristics: steady-state geothermal balance and transient thermal disturbance propagation, to comprehensively reflect the overall strength of geothermal coupling at the measuring points. The correlation assessment dimensions are more comprehensive and closely match the actual geothermal evolution law of the cavern.
[0056] Traditional multi-point weighted calculations generally employ linear distance attenuation, which does not conform to the objective physical law that underground geothermal heat decreases exponentially with increasing burial depth. The disturbance weights of shallow, distant measuring points cannot decay naturally, and intense shallow thermal noise excessively interferes with the calculation of deep geothermal trends. Therefore, a negative exponential distance attenuation factor can be constructed based on the spatial distance between different monitoring depths. Multiplying the basic synchronization degree by the distance attenuation factor yields the coupling weights of each monitoring depth for the corresponding component. The distance attenuation factor is a negative exponential function based on the spatial difference between the monitoring depth and the depth to be measured. Geothermal transfer naturally attenuates with increasing burial depth difference; the greater the distance, the weaker the influence of the measuring point on the geothermal activity at the depth to be measured.
[0057] The synchronous attenuation values at each monitoring depth are summed to obtain the total synchronous attenuation value. The ratio of the synchronous attenuation value at each monitoring depth to the total synchronous attenuation value is used as the weight of each depth, thus obtaining the long-term, periodic, and short-term weights for each monitoring depth.
[0058] In this embodiment, the long-term differential synchronization degree and the long-term component synchronization degree at each monitoring depth are added together and divided by 2 to obtain the long-term synchronization value at each monitoring depth. The periodic differential synchronization degree and the periodic component synchronization degree at each monitoring depth are added together and divided by 2 to obtain the periodic synchronization value at each monitoring depth. The short-term differential synchronization degree and the short-term component synchronization degree at each monitoring depth are added together and divided by 2 to obtain the short-term synchronization value at each monitoring depth. The long-term, periodic, and short-term synchronization values are multiplied by the distance attenuation factor to obtain the long-term, periodic, and short-term synchronization attenuation values, respectively.
[0059] The long-term synchronization attenuation values at each monitoring depth are summed to obtain the total long-term synchronization attenuation value. The ratio of the long-term synchronization attenuation value at each monitoring depth to the total long-term synchronization attenuation value is used as the weight for each depth, resulting in the long-term weight for each depth. Similarly, the periodic synchronization attenuation values at each monitoring depth are summed to obtain the total periodic synchronization attenuation value. The ratio of the periodic synchronization attenuation value at each monitoring depth to the total periodic synchronization attenuation value is used as the weight for each depth, resulting in the periodic weight for each depth. Finally, the short-term synchronization attenuation values at each monitoring depth are summed to obtain the total short-term synchronization attenuation value. The ratio of the short-term synchronization attenuation value at each monitoring depth to the total short-term synchronization attenuation value is used as the weight for each depth, resulting in the short-term weight for each depth. In this embodiment, the synchronization value at the deepest measurable depth is set to 1.
[0060] As an example, the distance attenuation factor satisfies the following formula: ; in, For the first Distance attenuation factor at each monitoring depth, It is a natural constant. For the first Each monitoring depth, For the depth to be predicted, The number for the monitoring depth, Take the absolute value. The distance adjustment parameter can be modified according to the thermal conductivity of the surrounding rock to correct the rate of attenuation. In this way, it conforms to the geothermal conduction attenuation characteristics of the strata, automatically reduces the contribution ratio of measuring points with large differences in burial depth from the depth to be measured, and naturally suppresses the interference of shallow long-distance instantaneous disturbances on deep geothermal benchmarks.
[0061] The coupling weight is a normalized weight that integrates the temporal geothermal synchronization characteristics and the dual constraints of formation spatial heat transfer attenuation. The coupling weight calculation method in this application accurately quantifies the spatial correlation between different measurable depths and the depth to be measured through a distance attenuation factor. The closer the distance, the greater the impact on the depth to be measured; the farther the distance, the smaller the impact. Simultaneously, it comprehensively characterizes the temperature correlation features between depths by combining the mean values of differential synchronization and component synchronization. By multiplying long-term, periodic, and short-term synchronization values by the distance attenuation factor, a dual coupling weighting of temperature characteristic correlation and spatial distance correlation is achieved, used to measure the impact on the depth to be measured at each depth.
[0062] The coupling weight of each component satisfies the following formula: ; in, For the first The monitoring depth is at the first The coupling weights of the components at time t, For the first The monitoring depth and the deepest measurable depth are at the following locations. The time-difference synchronization degree, For the first The monitoring depth and the deepest measurable depth are at the following locations. The degree of synchronization of components at any given moment. The number of monitoring depths is determined by the combined consideration of dynamic and static temporal correlations and the dual physical constraints of spatial heat transfer attenuation in the formation. The weight values are fully matched to the heat conduction logic of high geothermal strata, the contribution allocation of each measuring point is objective and reasonable, and the sum of the weights is normalized to ensure calculation stability.
[0063] Geothermal heat is conducted unidirectionally from deep to shallow, with the deepest point exhibiting the most stable geothermal signal. Shallow measuring points serve only as supplementary information, and equal weighting would dilute the core trend of deep natural geothermal activity. Therefore, a weighted summation of the corresponding components at all monitoring depths other than the measurable deepest point is performed using coupling weights. This summation is then superimposed on the corresponding component at the measurable deepest point to obtain the corrected long-term, periodic, and short-term components. The measurable deepest point represents the maximum burial depth at which sensors can reach within the cavern, where the geothermal signal is most stable. It serves as the reference point for geothermal activity throughout the cavern, and all shallow measuring point data are fused and corrected based on this point. The corrected components are new long-term / periodic / short-term components obtained by superimposing the weighted summation of all other measuring points onto the original component at the measurable deepest point, thus uniformly restoring the true geothermal evolution pattern of the entire geological strata.
[0064] The process of obtaining the long-term, periodic, and short-term correction components includes: multiplying the long-term weight at each depth by the long-term component at the corresponding monitoring depth, and summing the results of the multiplication at each monitoring depth to obtain the long-term correction component; multiplying the periodic weight at each depth by the periodic component at the corresponding depth, and summing the results of the multiplication at each monitoring depth to obtain the periodic correction component; and multiplying the short-term weight at each monitoring depth by the short-term component at the corresponding depth, and summing the results of the multiplication at each monitoring depth to obtain the short-term correction component.
[0065] The expressions for obtaining the long-term, periodic, and short-term correction components are all: ; ; in, These can be long-term correction components, periodic correction components, or short-term correction components. for The temperature value at time 1. for The Middle Temperature value at time, for The Middle Temperature value at time, For the first The depth-dependent long-term component, periodic component, or short-term component is at the th depth. Temperature value at time, For the length of time, The quantity at depth, For the first The depth is at the first Long-term, periodic, or short-term weighting of time points.
[0066] In this way, the original deep steady-state geothermal baseline can be preserved, and only the weighted correction of shallow measuring points can be superimposed to fully restore the true geothermal evolution characteristics of the entire strata. This solves the core defects of cross-depth correlation imbalance and deep trends being masked by shallow noise, and provides accurate subscale input data for subsequent feature extraction.
[0067] In step 104, the temperature values of adjacent moments in the corrected long-term component, periodic component, and short-term component are subtracted and the absolute values are taken to obtain the fluctuation values of each component. The fluctuation value of each component is the absolute value of the temperature difference between adjacent moments of a single component, quantifying the instantaneous temperature fluctuation amplitude of the component at the current moment and intuitively representing the strength of the instantaneous thermal disturbance. For a single corrected component (long-term / periodic / short-term), all time nodes are traversed, and the temperature of the previous moment is subtracted from the temperature of the next moment, and the absolute value of the difference is taken to generate the time series of the fluctuation values of the component throughout the entire time period. The three types of components are calculated independently, and each generates its own exclusive fluctuation value sequence. The temperature baseline offset interference is removed, and only the instantaneous fluctuation amplitude of each component is extracted to accurately quantify the dynamic fluctuation intensity of the formation geothermal, periodic conditions, and instantaneous disturbances at each moment, providing a pure disturbance intensity quantification index for dynamic weight calculation.
[0068] As an example, the fluctuation value satisfies the following formula: ; in, For the first Issue No. Time fluctuation value, For the first The first correction component in the period Temperature value at time, For the first The first correction component in the period Temperature value at time, This indicates that the first phase is long-term. This indicates that the second period is the cycle. This indicates that the third phase is short-term. It is an absolute value.
[0069] The total fluctuation value is obtained by summing the fluctuation values of each component at the same time. This total fluctuation value is the sum of the fluctuation values of the long-term, periodic, and short-term components at the same time, representing the overall temperature fluctuation of the cavern at that moment. By locking onto the same timestamp, the fluctuation values of the long-term, periodic, and short-term components are summed to obtain the total temperature fluctuation value for that moment. This process is repeated time-by-time to generate the total fluctuation time series. A unified global fluctuation reference benchmark is established for each moment, allowing for horizontal comparison of the fluctuation proportions of long-term geothermal, periodic, and instantaneous disturbance signals in the current period, enabling comparative and quantifiable weighting of different operating conditions over different time periods.
[0070] As an example, the total volatility value satisfies the following formula: ; in, For the first Total fluctuation value at any given time.
[0071] Next, the ratio of the fluctuation value of each component to the total fluctuation value is calculated to obtain the fluctuation contribution value of each component. The fluctuation contribution value is the ratio of the fluctuation value of a single component to the total fluctuation of all components, reflecting the dynamic proportion of each component solely based on the instantaneous fluctuation intensity. At any given moment, the fluctuation value of a single component ÷ the total fluctuation value = the fluctuation contribution value of that component. Long-term, periodic, and short-term components are calculated independently to obtain the proportion coefficient determined solely by the instantaneous fluctuation intensity. By simply quantifying the dynamic proportion of instantaneous disturbances, the most intense type of thermal disturbance activity in the current period is intuitively reflected, serving as the first-level calculation basis for dynamic weights.
[0072] As an example, the fluctuation contribution value satisfies the following formula: ; in, For the first Issue No. Contribution value of fluctuation over time.
[0073] Then, the fluctuation contribution value at the same moment is added to the preset basic contribution value to obtain the fused contribution value at the corresponding moment. The basic contribution value is a fixed constant set in advance based on the geothermal physical properties of high geothermal strata, used to characterize the inherent core status of long-term geothermal activity in the strata, and is not weakened by the magnitude of instantaneous fluctuations. A specific basic contribution constant for the long-term component (the basic contribution values for periodic and short-term components are lower) is set in advance based on high geothermal strata exploration data. The fluctuation contribution value at the same moment is superimposed with the corresponding category's basic contribution value to generate the fused contribution value. The fused contribution value is the sum of the fluctuation contribution value and the basic contribution value, taking into account both the intensity of instantaneous disturbances and the inherent geothermal basement of the strata. A specific inherent weight compensation for geothermal activity in the strata is introduced to compensate for the shortcoming of small fluctuations in the long-term component. Even during periods of stable geothermal activity without severe disturbances, the long-term geothermal component still has a basic weight, preventing the model from ignoring the evolution trend of the strata's background temperature.
[0074] As an example, the fusion contribution value satisfies the following formula: ; in, For the first Issue No. Contribution values are integrated at all times. For the first Issue No. Base contribution value at any given moment; Next, the fusion contribution values of the long-term, periodic, and short-term components are normalized to obtain the fluctuation contribution sequences of the long-term, periodic, and short-term components. The fluctuation contribution sequence is a time-series weight sequence obtained by normalizing the fusion contribution values across all time periods. This sequence serves as the adaptive weighting coefficient for the neural network features, with each time step corresponding to a set of weights specific to the three components. At the same time step, the fusion contribution values of the three components are summed to obtain the total. The normalized fluctuation contribution of that component is calculated by dividing the fusion contribution value of a single component by the total. This calculation is performed time-by-time to form a time-series weight sequence that dynamically changes with the operating conditions. At the same time step, the sum of the fluctuation contribution of the three components is always 1, and the weight values are standardized to ensure stable convergence of the neural network training. This generates time-by-time dynamically adjustable time-series weights, enabling adaptive allocation of feature importance based on the intensity of disturbances within the tunnel.
[0075] As an example, the volatility contribution sequence satisfies the following formula: ; in, For the first Issue No. The normalized fusion contribution values at each time point are arranged chronologically to obtain a long-term fluctuation contribution sequence. The normalized periodic fusion contribution values at each time point are also arranged chronologically to obtain a periodic fluctuation contribution sequence. Finally, the normalized short-term fusion contribution values at each time point are arranged chronologically to obtain a short-term fluctuation contribution sequence.
[0076] In this embodiment of the application, the formulas for obtaining the long-term, periodic, and short-term basic contribution values are as follows: In this embodiment of the application, a minimum value can be set in the denominator to avoid the denominator being 0.
[0077] Finally, the extracted features are weighted and fused based on the fluctuation contribution sequence to obtain fused temporal features. The fused temporal features are multi-scale temporal features output from three parallel channels. After element-wise weighting and fusion according to the fluctuation contribution at corresponding times, a unified feature vector is obtained and used as input to the prediction model. The three parallel channels output long-term, periodic, and short-term independent temporal features respectively. The fluctuation contribution of the three components at the same time is taken, and the corresponding channel features are scaled using element-wise multiplication. After weighting, all features are superimposed and merged to generate unified fused temporal features. In this way, the neural network can adaptively adjust the importance of the three types of features according to the real-time thermal disturbance state within the cave: appropriately reducing the weight of short-term noise during periods of severe disturbance and continuously increasing the weight of geothermal features during periods of stable geothermal conditions. This fully preserves multi-scale temporal details, weakens instantaneous disturbances with no engineering value, strengthens the core features of long-term geothermal evolution of the strata, and improves the stability of long-term deep temperature prediction.
[0078] By capturing the instantaneous fluctuation intensity of each correction component through the absolute value of the temperature difference between adjacent moments, and combining it with the basic contribution value to reflect the overall proportion of the component, the final fluctuation contribution sequence simultaneously takes into account both the "short-term fluctuation impact" and the "basic proportion impact" of the component on temperature changes, accurately depicting the dynamic contribution of different components to temperature changes. This step distinguishes between the instantaneous fluctuation intensity and the inherent geothermal base of the stratum to construct dynamic weights, overcoming the industry's shortcomings of simply allocating weights based on fluctuations and weakening long-term geothermal characteristics, and aligning with the temperature evolution essence of high-temperature underground caverns, which is "geothermal-dominated and disturbance-secondary." The proportion is calculated using the global total fluctuation as a unified reference, coupled with dedicated geothermal basic contribution compensation, and stable temporal weights are generated after normalization, resulting in better convergence of neural network training and unified feature fusion scale. It achieves full-time adaptive feature weighting, automatically suppressing short-term noise weights during periods of severe disturbance such as blasting and ventilation inside the cavern, and continuously focusing on the formation temperature evolution law during long-term stable geothermal stages, ensuring that multi-scale features do not mask each other, and significantly improving the accuracy of long-term temperature prediction for deep locations without monitoring points.
[0079] In step 105, the absolute temperature upper limit threshold and the allowable temperature rise rate threshold from the safety specifications for construction of high-temperature underground caverns are first obtained. The absolute temperature upper limit threshold is the maximum allowable temperature limit of the stratum defined in the safety specifications for construction of deep high-temperature underground caverns. It is determined comprehensively by the heat resistance characteristics of the surrounding rock, the safety of construction personnel, and the high-temperature resistance performance of the lining support. It represents the critical temperature value that does not produce geological heat hazards and ensures construction safety after the superposition of natural geothermal heat and anthropogenic heat. The allowable temperature rise rate threshold is the maximum allowable temperature rise per unit time specified in the safety specifications for high-temperature underground caverns. Rapid temperature rise of the stratum can easily induce thermal cracking of the surrounding rock and thermal damage to the support structure, and is a core dynamic indicator for predicting thermal disasters. The special construction safety specifications for deep high-temperature underground engineering and the design standards for geothermal protection of surrounding rock are retrieved to extract two types of fixed safety constraint parameters: one is the absolute temperature upper limit of the entire stratum; the other is the maximum allowable temperature rise per unit time. The two sets of thresholds are solidified and stored in the system verification module as the judgment benchmark. By introducing mandatory safety constraints specific to engineering projects and establishing standardized geothermal safety judgment benchmarks, we can provide clear and quantifiable evaluation criteria for the compliance verification of prediction results, and ensure that the model output results meet the mandatory requirements for construction control of high geothermal caverns.
[0080] Next, the rate of change of the initial predicted temperature relative to the predicted temperature at the previous moment is calculated. The initial predicted temperature is a pure data prediction value directly output by the model integrating time-series features, without undergoing geothermal safety compliance verification. It is obtained solely based on time-series data fitting and does not incorporate formation or construction safety constraints. The rate of temperature change is the ratio of the temperature difference between two adjacent prediction moments to the time interval, characterizing the real-time speed of formation temperature rise and used to identify sudden thermal anomalies. The initial predicted temperature at the current moment and the model output temperature at the previous moment are read, and the difference between the two is divided by the time sampling interval to obtain the real-time temperature change rate, quantifying the short-term temperature rise speed. In this way, both static temperature amplitude and dynamic temperature rise trend risk dimensions can be covered simultaneously, capturing the precursors of short-term rapid temperature rises and compensating for the shortcomings of single static threshold judgment, thus improving the early warning of thermal hazards.
[0081] Next, it is determined whether the initial predicted temperature exceeds the absolute temperature upper limit threshold, or whether the rate of change exceeds the allowable heating rate threshold. If either the instantaneous temperature exceeds the upper limit or the heating rate exceeds the limit, it is determined to be a geothermal anomaly. Only when both indicators meet the limit requirements is it considered compliant. A strict double OR logic is used for judgment; if any safety indicator exceeds the standard limit, an anomaly is identified, eliminating the risk of missed detection and improving the completeness of high-temperature hazard identification.
[0082] If the initial predicted temperature is not greater than the upper limit threshold of absolute temperature and the rate of change is not greater than the allowable heating rate threshold, then the initial predicted temperature will be output as the target predicted temperature. The target predicted temperature is a final usable predicted value that has been verified and corrected by dual geothermal safety thresholds and simultaneously conforms to the objective laws of geothermal formation and construction safety requirements. It can be directly used for cavern geothermal management and construction plan adjustments. In compliant scenarios, it directly outputs accurate geothermal prediction values from the model, fully preserves the real geothermal evolution data of the formation, and automatically marks the safety status without manual verification, making it suitable for routine cavern geothermal monitoring.
[0083] If the initial predicted temperature exceeds the absolute temperature upper limit threshold, or the rate of change exceeds the allowable temperature rise rate threshold, a high-temperature warning signal is triggered, and the target predicted temperature is forcibly corrected to the absolute temperature upper limit threshold before being output. The high-temperature warning signal is an alarm command linked to the system's monitoring platform and on-site control terminals, used to remind maintenance personnel to handle high ground temperature hazards and adjust ventilation and cooling measures. When geothermal anomalies are detected, the system pushes high-temperature warning commands to the monitoring platform and on-site control equipment. Simultaneously, the output target predicted temperature is forcibly locked to the absolute temperature upper limit specified in the standards, and no longer outputs out-of-limit fitted values. False out-of-limit predicted values are forcibly corrected, ensuring the output value conforms to the physical limits of the strata, avoiding misleading on-site ventilation and support control. The automatic linkage alarm mechanism achieves an integrated closed loop of prediction, risk identification, and early warning, eliminating the need for real-time manual monitoring and allowing for proactive handling of high ground temperature hazards. Furthermore, it balances the dual requirements of data prediction accuracy and construction safety management, significantly improving the practical value of the solution for engineering implementation.
[0084] Figure 2 This is a schematic diagram of the structure of a temperature prediction neural network provided in an embodiment of this application. Figure 2 As shown, the temperature prediction neural network includes: multiple temporal feature extraction channels, a multi-component feature fusion layer, a first LSTM layer, a Concat layer, a first fully connected layer, and a second fully connected layer.
[0085] The first time-series feature extraction channel is used to extract time-series features from the long-term correction component, resulting in long-term time-series features. The second time-series feature extraction channel is used to extract time-series features from the periodic correction component, resulting in periodic time-series features. The third time-series feature extraction channel is used to extract time-series features from the short-term correction component, resulting in short-term time-series features.
[0086] The multi-component feature fusion layer is used to fuse long-term, periodic, and short-term time-series features based on the contribution sequences of long-term, periodic, and short-term fluctuations, resulting in multi-component fused features. The first LSTM layer extracts time-series fused features from the multi-component fused features. The first fully connected layer maps the depth to be measured into depth features. The Concat layer concatenates the depth features and the time-series fused features to obtain concatenated features. The second fully connected layer uses the concatenated features to obtain the predicted temperature of the depth to be measured.
[0087] The temperature prediction neural network in this application extracts temporal features from long-term, periodic, and short-term correction components through channels. Simultaneously, it guides the fusion of multi-component features based on the fluctuation contribution sequence, strengthening the weight of key features. Then, it combines the depth features of the measured depth with the temporal fusion features to perform temperature prediction.
[0088] In this embodiment of the application, the depth to be measured of the first fully connected layer is a normalized value.
[0089] In this embodiment, the expression for the multi-component feature fusion layer is: ; in, For multi-component fusion features, For the long-term fluctuation contribution sequence, For the contribution sequence of periodic fluctuations, As a contribution sequence to short-term fluctuations, For long-term time series characteristics, It is a periodic time series feature. It is a short-term time series characteristic. This is element-wise multiplication.
[0090] In this embodiment, the fluctuation contribution sequence is obtained by combining the fluctuation contribution value and the basic contribution value. Therefore, , , When weighting, the focus is on the characteristics corresponding to components with high fluctuation and high temperature values. The basic contribution value reflects the overall proportion of the component in terms of temperature value, while the fluctuation contribution value reflects the instantaneous change intensity of the component. This achieves dual weighting guidance of "static proportion + dynamic fluctuation": when a component has both high temperature proportion and strong fluctuation characteristics, its corresponding time series characteristics will be significantly enhanced in the fusion result, so that the fusion characteristics can take into account both the "long-term dominant component" and the "short-term key fluctuation value" of temperature change.
[0091] In the embodiments of this application, the long-term correction component, the periodic correction component, the short-term correction component, the long-term fluctuation contribution sequence, the periodic fluctuation contribution sequence, and the short-term fluctuation contribution sequence are aligned in time and maintain consistent length.
[0092] Figure 3 This is a schematic diagram of the structure of multiple temporal feature extraction channels provided in the embodiments of this application. For example... Figure 3 As shown, the multiple temporal feature extraction channels have the same structure, all of which include the following sequentially connected layers: a first one-dimensional convolutional layer, a second one-dimensional convolutional layer, a third one-dimensional convolutional layer, and a second LSTM layer.
[0093] In this embodiment, the kernel size of the first one-dimensional convolutional layer, the second one-dimensional convolutional layer, and the third one-dimensional convolutional layer are all set to 3, the stride is 1, and the padding is 1, so that the input data and output data of the temporal feature extraction channel have the same length.
[0094] This application utilizes EMD to decompose historical temperature sequences at different depths into three independent components: long-term trend, periodic variation, and short-term fluctuation, achieving precise separation of different feature dimensions of the temperature signal. Compared to the method of directly modeling the overall features of the temperature sequence using the tower-breaking technique, this method can effectively avoid interference between different feature components and accurately capture the independent contribution of each component to the temperature changes in underground caverns.
[0095] This application quantifies the correlation between different depths and the deepest measurable depth in various feature components from two dimensions: temporal differential synchronization and component synchronization. It also constructs long-term, periodic, and short-term weights for each depth using a distance attenuation factor. Temporal differential synchronization focuses on the synchronicity of component changes at adjacent times, component synchronization reflects the overall matching degree of the components themselves, and the distance attenuation factor conforms to the law of attenuation of the influence of underground stratum temperature with depth. This application uses long-term, periodic, and short-term weights at depth to weight and correct the corresponding components, ensuring that the temperature components at each depth reflect the true influence relationship of the depth being measured. This avoids the cross-depth correlation imbalance problem caused by directly using the original components, considers the impact of long-term trends, periodic changes, and short-term fluctuations in the temperature signal on the temperature correlation at different depths, and improves the accuracy of temperature prediction.
[0096] This application embodiment employs a temperature prediction neural network to jointly process long-term, periodic, and short-term correction components, and combines the depth to be measured with feature weighting based on the fluctuation contribution sequences of long-term, periodic, and short-term components. This enables the neural network to adaptively focus on features with large correction components and strong fluctuations, thereby improving the accuracy of temperature prediction.
[0097] Figure 4 This is a schematic diagram of the structure of a temperature prediction device for a high-temperature underground cavern complex provided in an embodiment of this application. Figure 4 As shown, the temperature prediction device 400 for the high geothermal underground cavern complex may include an acquisition module 401, a decomposition module 402, a correction module 403, an extraction module 404, and an output module 405.
[0098] The acquisition module 401 is used to acquire historical temperature sequences at different monitoring depths of high-temperature underground caverns, and to preprocess the historical temperature sequences in combination with geological prior information, including rock thermal conductivity and geothermal gradient.
[0099] The decomposition module 402 is used to decompose the preprocessed historical temperature sequence into long-term components, periodic components, and short-term components using the empirical mode decomposition method.
[0100] The correction module 403 is used to calculate the coupling weight of each component in the depth direction based on the spatial position relationship and thermal conduction characteristics between different monitoring depths, and to use the coupling weight to perform weighted correction on the corresponding components at each depth to obtain the corrected long-term component, periodic component and short-term component.
[0101] The extraction module 404 is used to input the corrected long-term component, the periodic component and the short-term component into three parallel time-series feature extraction channels for feature extraction, and to perform weighted fusion of the extracted features based on the fluctuation contribution of each component to obtain fused time-series features.
[0102] The output module 405 is used to input the fused time series features into the prediction model, output the initial predicted temperature of the depth to be measured, verify the initial predicted temperature through geothermal compliance rules, and output the final target predicted temperature.
[0103] The acquisition module 401, decomposition module 402, correction module 403, extraction module 404 and output module 405 can be used to execute steps 101-105 in the embodiments of the above-mentioned high geothermal underground cavern group temperature prediction method. For the specific implementation of these modules and more details, please refer to the corresponding method section, which will not be elaborated here.
[0104] This application also provides a computer-readable storage medium storing a program that can be loaded by a processor and executed by any of the temperature prediction methods for high-temperature underground cavern groups in this application.
[0105] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.
[0106] The above examples illustrate this application only to aid understanding and are not intended to limit its scope. Those skilled in the art to which this application pertains can make various simple deductions, modifications, or substitutions based on the ideas presented.
Claims
1. A method for predicting the temperature of a high-temperature underground cavern complex, characterized in that, include: Historical temperature sequences at different monitoring depths of high-temperature underground caverns are obtained, and the historical temperature sequences are preprocessed in combination with prior geological information, including rock thermal conductivity and geothermal gradient. The preprocessed historical temperature sequence is decomposed into long-term components, periodic components, and short-term components using the empirical mode decomposition method. Based on the spatial relationship and thermal conduction characteristics between different monitoring depths, the coupling weight of each component in the depth direction is calculated, and the corresponding components at each depth are weighted and corrected using the coupling weight to obtain the corrected long-term component, periodic component and short-term component. The corrected long-term component, periodic component, and short-term component are respectively input into three parallel time-series feature extraction channels for feature extraction. The extracted features are then weighted and fused based on the fluctuation contribution of each component to obtain fused time-series features. The fused temporal features are input into the prediction model, the initial predicted temperature of the depth to be measured is output, and the initial predicted temperature is verified by the geothermal compliance rules to output the final target predicted temperature.
2. The temperature prediction method for high-temperature underground cavern groups according to claim 1, characterized in that, The preprocessing of the historical temperature sequence by incorporating prior geological information includes: Based on the geothermal gradient in the geological prior information, a theoretical temperature reference curve for each monitoring depth is constructed, and the deviation between the temperature value at each moment in the historical temperature sequence and the theoretical temperature reference curve is calculated. The deviation value is used as an effective temperature sequence to characterize local thermal anomaly disturbances. Based on the thermal conductivity of the rock mass in the geological prior information, the thermal conduction time constant of the effective temperature sequence at different rock layer depths is determined. The effective temperature sequence is then subjected to hysteresis compensation processing using the thermal conduction time constant to obtain the preprocessed historical temperature sequence after eliminating spatial thermal conduction delay.
3. The temperature prediction method for high-temperature underground cavern groups according to claim 1, characterized in that, The step of decomposing the preprocessed historical temperature sequence into long-term, periodic, and short-term components using the empirical mode decomposition method includes: Based on the local extreme points of the historical temperature sequence, upper and lower envelopes are constructed by cubic spline interpolation and local mean is calculated. Local mean is eliminated through an iterative sieving process until the preset intrinsic mode function stopping criterion is met, resulting in multiple intrinsic mode functions and residual components arranged in descending order of frequency. The first eigenmode function with the highest frequency is extracted as the short-term component characterizing the transient thermal disturbance. The residual component is extracted as the long-term component characterizing the slow evolution trend of the deep original geothermal field; The remaining intrinsic mode functions, excluding the first intrinsic mode function, are numerically superimposed at the same time point to reconstruct the periodic component that characterizes the periodic impact of the cavern environment and construction activities.
4. The temperature prediction method for high-temperature underground cavern groups according to claim 1, characterized in that, The calculation of the coupling weight of each component in the depth direction based on the spatial positional relationship and thermal conduction characteristics between different monitoring depths includes: For each component at each monitoring depth, the temperature values at adjacent times are subtracted to obtain a time-series difference sequence; The Hilbert transform is performed on the time-series differential sequence to extract the differential instantaneous phase. At the same time, the differential instantaneous phase difference between other monitoring depths and the deepest measurable depth is calculated to obtain the differential synchronization degree characterizing the consistency of the transient thermal disturbance propagation trend. For each component at each monitoring depth, Hilbert transform is directly performed to extract the component's instantaneous phase. At the same time, the instantaneous phase difference between other monitoring depths and the deepest measurable depth is calculated to obtain the component synchronization degree characterizing the steady-state thermal equilibrium correlation. The basic synchronization degree is obtained by averaging the differential synchronization degree and the component synchronization degree.
5. The temperature prediction method for high-temperature underground cavern groups according to claim 4, characterized in that, The step of using the coupling weights to perform weighted correction on the corresponding components at each depth includes: Based on the spatial distance between different monitoring depths, a distance attenuation factor in the form of a negative exponential form is constructed. Multiplying the basic synchronization degree by the distance attenuation factor yields the coupling weight of each monitoring depth on the corresponding component; The corresponding components of other monitoring depths besides the measurable deepest depth are weighted and summed using the coupling weights, and the summation results are superimposed on the corresponding component of the measurable deepest depth to obtain the corrected long-term component, the periodic component, and the short-term component.
6. The temperature prediction method for high-temperature underground cavern groups according to claim 5, characterized in that, The distance attenuation factor satisfies the following formula: ; in, For the first Distance attenuation factor at each monitoring depth, It is a natural constant. For the first Each monitoring depth, For the depth to be predicted, The number for the monitoring depth, Take the absolute value. Adjust the distance parameters; The coupling weight of each component satisfies the following formula: ; in, For the first The monitoring depth is at the first The coupling weights of the components at time t, For the first The monitoring depth and the deepest measurable depth are at the following locations. The time-difference synchronization degree, For the first The monitoring depth and the deepest measurable depth are at the following locations. The degree of synchronization of components at any given moment. The number of depths to be monitored.
7. The temperature prediction method for high-temperature underground cavern groups according to claim 1, characterized in that, The extracted features are weighted and fused based on the fluctuation contribution of each component to obtain fused temporal features, including: The temperature values at adjacent moments in the corrected long-term component, periodic component, and short-term component are subtracted and the absolute values are taken to obtain the fluctuation values of each component. The total fluctuation value is obtained by summing the fluctuation values of each component at the same time. Calculate the ratio of the fluctuation value of each component to the total fluctuation value to obtain the fluctuation contribution value of each component. The fluctuation contribution value at the same moment is added to the preset basic contribution value to obtain the fusion contribution value at the corresponding moment; The fusion contribution values of the long-term component, the periodic component, and the short-term component are normalized respectively to obtain the fluctuation contribution sequence of the long-term component, the periodic component, and the short-term component. The extracted features are weighted and fused based on the fluctuation contribution sequence to obtain fused temporal features.
8. The temperature prediction method for high-temperature underground cavern groups according to claim 1, characterized in that, The process of verifying the initial predicted temperature using geothermal compliance rules and outputting the final target predicted temperature includes: Obtain the absolute temperature upper limit threshold and the allowable heating rate threshold in the safety specifications for the construction of high geothermal caverns; Calculate the rate of change of the initial predicted temperature relative to the predicted temperature at the previous time step; Determine whether the initial predicted temperature is greater than the absolute temperature upper limit threshold, or determine whether the rate of change is greater than the allowable heating rate threshold; If the initial predicted temperature is not greater than the absolute temperature upper limit threshold and the rate of change is not greater than the allowable heating rate threshold, then the initial predicted temperature is output as the target predicted temperature. If the initial predicted temperature is greater than the absolute temperature upper limit threshold, or the rate of change is greater than the allowable heating rate threshold, a high temperature warning signal is triggered, and the target predicted temperature is forcibly corrected to the absolute temperature upper limit threshold before being output.
9. A temperature prediction device for a high-temperature underground cavern complex, characterized in that, include: The acquisition module is used to acquire historical temperature sequences at different monitoring depths of high-temperature underground caverns, and to preprocess the historical temperature sequences in combination with prior geological information, including rock thermal conductivity and geothermal gradient. The decomposition module is used to decompose the preprocessed historical temperature sequence into long-term components, periodic components, and short-term components using the empirical mode decomposition method. The correction module is used to calculate the coupling weight of each component in the depth direction based on the spatial position relationship and heat conduction characteristics between different monitoring depths, and to use the coupling weight to perform weighted correction on the corresponding components at each depth to obtain the corrected long-term component, periodic component and short-term component. The extraction module is used to input the corrected long-term component, the periodic component and the short-term component into three parallel time-series feature extraction channels for feature extraction, and to perform weighted fusion of the extracted features based on the fluctuation contribution of each component to obtain fused time-series features. The output module is used to input the fused temporal features into the prediction model, output the initial predicted temperature of the depth to be measured, verify the initial predicted temperature through geothermal compliance rules, and output the final target predicted temperature.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that can be loaded by a processor and executed as described in any one of claims 1 to 8, the method for predicting the temperature of high-temperature underground cavern groups.
