Geothermal resource uncertainty quantification assessment method, system, device and medium

Abstract

The present application belongs to the technical field of geothermal resource exploration and evaluation, and discloses a geothermal resource uncertainty quantification evaluation method, system, device and medium. The method comprises the following steps: extracting surface temperature anomaly from thermal infrared image; taking the anomaly, fault distance and lithospheric thickness as features, obtaining predicted heat flow grid and statistical uncertainty through regularization regression; constructing a multi-layer lithospheric thermal parameter structure, and obtaining lithology parameter uncertainty through parameter disturbance forward; constructing a physical information neural network with a discard layer, and obtaining model uncertainty through multiple forward propagation; merging the three kinds of uncertainty into comprehensive uncertainty, and performing resource target classification based on the comprehensive uncertainty, adaptive reliability distance threshold and favorable degree score after percentile truncation normalization. The present application realizes the collaborative fusion of remote sensing surface data, scattered point measurement and physical constraint, can constrain the geothermal resource classification by the whole link uncertainty, and is suitable for regional to exploration scale geothermal target area decision support.

Description

Technical Field

[0001] This invention relates to the field of geothermal resource exploration and evaluation technology, and in particular to a quantitative assessment method, system, equipment and medium for the measurement of geothermal resource uncertainty. Background Technology

[0002] The core tasks of geothermal resource assessment include identifying favorable target areas and quantifying resource quantities. Decision-makers use this information to determine drilling locations and priorities. The level of quantification of the uncertainty in predicting heat flow is directly related to the risk control capability of exploration investment.

[0003] Geothermal target area evaluation methods can be divided into three categories: traditional statistical regression based on geological-geophysical data, thermal anomaly delineation based on satellite surface temperature, and physical-data hybrid modeling methods (such as Physical Information Neural Network, PINN). Each of these three methods has its strengths in heat flow prediction accuracy, but all suffer from the following shortcomings in uncertainty quantification: First, the uncertainty of lithological thermal parameters is severely underestimated. From the perspective of thermal property modeling, the thermal structure of the lithosphere is usually simplified to four layers: sedimentary cover, upper crustal crystalline basement, lower crustal granulite facies layer, and lithospheric mantle. The thermal conductivity of each layer can differ by more than two times, and the heat production rate can differ by an order of magnitude. Laboratory measurements show that the relative standard deviations of rock thermal conductivity and heat production rate are approximately 15% and 25%, respectively. When these parameter deviations are mapped to surface heat flow through forward modeling of heat conduction, they exhibit nonlinear differential propagation due to the different proportions of thermal resistance in each layer: the thermal resistance proportion of shallow sedimentary layers is significantly higher than their thickness proportion, and their parameter deviations have a correspondingly higher weight on surface heat flow; while the thermal resistance proportion of deep mantle layers is lower, and the influence of their parameter deviations is attenuated. If the lithosphere is simplified to a homogeneous single layer, the sensitivity of the equivalent thermal conductivity is globally averaged, leading to an underestimation of parameter uncertainty. Fuchs et al. (2018) showed in their study of the Danish Basin that the uncertainty introduced by using the homogeneous thermal conductivity values ​​of the whole region for temperature modeling can reach an average of 20%, and is even higher in some local areas.

[0004] Second, the Physical Information Neural Network (PINN) output lacks model structural uncertainty. Existing geothermal physical information neural networks (PINNs) only output a deterministic temperature field and a single predicted heat flux after training. Because PINN training involves non-convex weighted optimization of data fitting and physical residual terms, the network weights typically converge to a local minimum of the loss function, thus fixing the prediction and failing to express the crucial information that "the model's confidence decreases in extrapolated regions far from observation constraints." This lack of model uncertainty makes subsequent resource classification unable to distinguish between high-confidence predictions in sufficiently constrained regions and low-confidence inferences in extrapolated regions. Artificially high heat fluxes far from observation points may be incorrectly classified as high-level target areas.

[0005] Third, the reliability distance threshold relies on fixed empirical values, which cannot adapt to different observation densities. Existing systems mostly use preset fixed thresholds based on Euclidean distance or geological distance from the nearest measuring point to divide high, medium, and low reliability zones. When stations are densely packed in the study area, the near-end threshold is too lenient, causing remote areas to receive undeservedly high reliability scores; when stations are sparse, the same threshold is too strict, causing most areas to fall into the low reliability level and lose their guiding significance. A reasonable distance classification should be linked to the actual spatial density of observations. When observation points are denser, the near-end high-confidence radius should be narrower; when observation points are sparser, the distance classification transition should be appropriately relaxed while retaining a higher uncertainty indicator.

[0006] Fourth, the favorability score normalization lacks robustness to extreme outliers. Thermal assessments typically normalize and weight multiple sources of evidence, such as heat flow anomalies, surface temperature anomalies, fault proximity, and lithospheric thinning, to synthesize a comprehensive favorability score. Existing normalization methods generally employ min-maximum linear scaling, using the minimum and maximum values ​​of the data as mapping endpoints. This method is highly sensitive to extreme outliers: in active geothermal systems, the surface temperature of a single crater pixel can reach 15 times the background mean, causing more than 98% of normal pixels to be compressed into an extremely narrow range. This results in the favorability score losing its spatial discriminative power and failing to distinguish between gradient changes in moderate and weak anomalies within normal regions.

[0007] The above four shortcomings are interconnected: the underestimation of lithological parameter uncertainty makes the confidence band of the statistical path too narrow; the lack of uncertainty in the PINN model makes it unable to provide a prediction confidence interval; the fixed distance threshold makes it impossible to uniformly measure and connect the uncertainty information of different methods in space; and the fragility of normalization introduces additional systematic bias in the final favorable outcome synthesis. Summary of the Invention

[0008] To address the aforementioned issues, this invention provides a method, system, device, and medium for quantitative assessment of geothermal resource uncertainty. It resolves problems in existing technologies such as underestimated lithological thermal parameter uncertainty, lack of model structure uncertainty in physical information neural networks, reliance on fixed empirical values ​​for reliability distance thresholds that cannot adapt to different observation densities, and lack of robustness of advantage score normalization to extreme outliers. The invention employs multi-layered thermal resistance Monte Carlo forward modeling to accurately propagate layered parameter uncertainties, discards inference in physical information neural networks to obtain model structure uncertainty, adaptively calculates reliability distance thresholds based on the median nearest neighbor distance of observation points, and uses percentile truncation normalization to suppress the compression effect of extreme outliers on the score distribution. This achieves end-to-end uncertainty quantification and reliable resource target differentiation from statistical regression and lithological parameters to model structure.

[0009] According to a first technical solution of the present invention, a quantitative assessment method for geothermal resource uncertainty is provided, the method comprising: Extracting surface temperature anomalies in the target area from thermal infrared images; Using the surface temperature anomaly, fault distance, and lithosphere thickness as feature matrices, and measured geothermal heat flow observation points as point source constraints, a predicted heat flow grid and its statistical regression uncertainty are obtained through regularized regression. The predicted heat flow grid is then cropped to a set physical range, and pixels exceeding the boundary are marked with reliability degradation. Uncertainty is propagated according to preset rules to obtain the statistical uncertainty after physical range constraints. Based on the layered interface data of the target region, a multi-layered lithosphere thermal parameter structure is constructed. By applying perturbation to the thermal conductivity and heat generation rate of each layer and performing multi-layered heat conduction forward modeling and repeated calculations, the uncertainty of lithological parameters is obtained. A physical information neural network is constructed with spatial coordinates as input and temperature as output. Each hidden layer of the physical information neural network contains a dropout layer. During training, the loss function incorporates the following constraints: remotely sensed surface temperature Dirichlet constraint corrected for shallow bias, measured heat flux Neumann flux constraint, borehole temperature internal point constraint, variable-depth lithosphere bottom interface temperature constraint, layered steady-state heat conduction physical residual term, and soft constraint that temperature monotonically increases with depth. After training, the dropout layer is kept active and multiple forward propagations are performed. The model uncertainty is obtained based on the pixel-by-pixel standard deviation of the multiple forward propagations. The statistical uncertainty, lithological parameter uncertainty, and model uncertainty are combined into a pixel-level comprehensive uncertainty. Based on the comprehensive uncertainty, the adaptive reliability distance threshold, and the percentile truncation normalized advantage score, resource target classification is performed on the target area.

[0010] Furthermore, the statistical uncertainty after physical range constraints is calculated using the following formula. : in, This is the lower limit of observation error. This is the penalty coefficient that increases with the distance of the boundary crossing. For pixel-level statistical regression uncertainty, It is a function for maximizing the value; The penalty coefficient Represented as: in, The penalty intensity coefficient, The predicted surface heat flow value, This represents the physical upper limit of surface heat flow. This represents the physical lower limit of surface heat flow.

[0011] Furthermore, a multi-layered lithospheric thermal parameter structure is constructed based on the layered interface data of the target region. By perturbing the thermal conductivity and heat production rate of each layer and performing multi-layered heat conduction forward modeling repetitive calculations, the lithological parameter uncertainties are obtained, including: for Layered lithosphere, obtaining the thickness of each layer Thermal conductivity and volumetric heat production rate Based on the current predicted heat flow and multilayer thermal parameters, the equivalent top-to-bottom temperature difference of the lithosphere can be inversely calculated using the following formula. : in, As surface heat flow, The index represents the number of layers in the lithosphere; the above formula is an approximation of the equivalent thermal resistance under one-dimensional steady-state heat conduction conditions, used for parameter perturbation sensitivity estimation.

[0012] After perturbing the thermal conductivity and heat generation rate of each layer, the equivalent top-bottom temperature difference... Under the constraints, the disturbed surface heat flow is calculated using the following formula. : in, For the perturbation of the first Layer thermal conductivity, For the perturbation of the first Layer volumetric heat production rate; Take the result from multiple calculations The pixel-by-pixel standard deviation as the uncertainty of lithological parameters .

[0013] Furthermore, the formula for determining the comprehensive uncertainty is: in, To account for the overall uncertainty, The statistical uncertainty is the result of physical range constraints. For the uncertainty of lithological parameters, This represents the model uncertainty.

[0014] Furthermore, in the process of extracting surface temperature anomalies of the target area from thermal infrared images, when soil or shallow cover thermal resistance data are available, the surface temperature anomalies are converted into equivalent heat flux indicators after surface thermal resistance correction using the following formula: in, This is an equivalent heat flux indicator. This is due to abnormal surface temperature. This represents the surface thermal resistance.

[0015] Furthermore, the shallow bias-corrected remote sensing surface temperature surface Dirichlet constraint uses the extracted surface temperature anomaly as the upper boundary condition. The bottom interface depth function used for the temperature constraint of the variable-depth lithosphere bottom interface is obtained by spatial interpolation from the lithosphere thickness grid. Bilinear interpolation is used for regular grids, while it degenerates into inverse distance weighted or triangular mesh linear interpolation for irregular meshes. The soft constraint, in which temperature monotonically increases with depth, is added to the total loss with an adjustable weighting coefficient and only applies to collocation points below a preset shallow depth.

[0016] Furthermore, based on the comprehensive uncertainty, adaptive reliability distance threshold, and percentile-truncated normalized advantage score, resource target classification and grading of the target region are performed, including: The adaptive reliability distance thresholds are dynamically calculated based on the median nearest neighbor distance of the measured heat flow observation points, and are used as the near-end, middle-end, and far-end reliability distance thresholds. The percentile truncation normalization is obtained by truncating and normalizing each item of the advantage score using preset low percentile thresholds and high percentile thresholds. Based on the comprehensive uncertainty, the reliability distance threshold level to which the distance from each pixel to the nearest observation point belongs, and the favorable score after percentile truncation and normalization, the target area is divided into multiple resource levels, among which pixels with comprehensive uncertainty exceeding the preset threshold are downgraded.

[0017] According to a second technical solution of the present invention, a quantitative assessment system for geothermal resource uncertainty is provided, the system comprising: The data extraction module is configured to extract surface temperature anomalies of the target area from thermal infrared images; The first uncertainty calculation module is configured to use the surface temperature anomaly, fault distance, and lithosphere thickness as feature matrices, and measured geothermal heat flow observation points as point source constraints. It obtains the predicted heat flow grid and its statistical regression uncertainty through regularized regression. The predicted heat flow grid is then clipped to a set physical range, and pixels exceeding the boundary are marked with reliability degradation. Uncertainty is propagated according to preset rules to obtain the statistical uncertainty after physical range constraints. The second uncertainty calculation module is configured to construct a multi-layered lithosphere thermal parameter structure based on the layered interface data of the target region. By applying perturbation to the thermal conductivity and heat generation rate of each layer and performing multi-layered heat conduction forward modeling repeated calculations, the uncertainty of lithological parameters is obtained. The third uncertainty calculation module is configured to construct a physical information neural network with spatial coordinates as input and temperature as output. Each hidden layer of the physical information neural network contains a dropout layer. During training, the loss function incorporates the following constraints: remotely sensed surface temperature Dirichlet constraint corrected for shallow bias, measured heat flux Neumann flux constraint, well temperature internal point constraint, variable-depth lithosphere bottom interface temperature constraint, layered steady-state heat conduction physical residual term, and soft constraint that temperature monotonically increases with depth. After training, the dropout layer is kept active and multiple forward propagations are performed. The model uncertainty is obtained based on the pixel-by-pixel standard deviation of the multiple forward propagations. The target classification module is configured to merge the statistical uncertainty, lithological parameter uncertainty and model uncertainty into a pixel-level comprehensive uncertainty, and classify the target area based on the comprehensive uncertainty, the adaptive reliability distance threshold and the percentile truncation normalized advantage score.

[0018] According to a third technical solution of the present invention, an electronic device is provided, the electronic device comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the method described above.

[0019] According to a fourth technical solution of the present invention, a non-transitory computer-readable storage medium storing instructions is provided, which, when executed by a processor, performs the method described above.

[0020] The geothermal resource uncertainty quantitative assessment methods, systems, equipment, and media according to various schemes of the present invention have at least the following technical effects: First, this invention employs a multilayer thermal resistance forward model, incorporating thermal conductivity into the calculation in the form of thermal resistance. This ensures that the shallow, low-thermal-conductivity layer, due to its larger proportion of thermal resistance, naturally has a higher weight in influencing surface heat flow than the deep, high-thermal-conductivity layer. This accurately reflects the differentiated contribution of the layered structure to uncertainty propagation and avoids the underestimation of parameter uncertainty caused by simplifying the lithosphere into a homogeneous single layer.

[0021] Second, this invention sets a dropout layer after each hidden layer of the physical information neural network, and keeps the dropout layer active during the inference phase, performing multiple forward propagations to obtain the model structural uncertainty through pixel-by-pixel standard deviation. This Monte Carlo dropout inference method can efficiently obtain model uncertainty without training multiple independent network models, and the uncertainty remains at a low level in regions with sufficient observation constraints, naturally increasing to 3 to 4 times that of the constrained region in extrapolation regions far from observation, which is consistent with the theoretical expectation of Bayesian inference.

[0022] Third, this invention dynamically calculates the reliability distance thresholds for near-end, middle-end, and far-end based on the median nearest neighbor distance of measured heat flow observation points, so that the reliability judgment thresholds are adaptively adjusted with the actual observation density: the high confidence radius of the near end is automatically narrowed in densely populated areas of stations, making the reliability judgment more stringent; the grade transition is appropriately relaxed in sparsely populated areas of stations, maintaining a high uncertainty indicator while retaining gradient information, effectively solving the problem of poor adaptability of fixed empirical thresholds.

[0023] Fourth, this invention uses preset low percentile and high percentile thresholds to truncate and normalize each item of the favorability score, effectively suppressing the compression effect of extreme outliers on the score distribution. Compared with the linear scaling of minimum and maximum values ​​used in the prior art, the percentile truncation normalization of this invention significantly improves the spatial discrimination of the favorability score, and clearly distinguishes the gradual characteristics of moderate and weak outliers within the normal region.

[0024] Fifth, this invention combines the statistical uncertainty after physical range constraints, the lithological parameter uncertainty obtained from multi-level Monte Carlo forward modeling, and the model uncertainty obtained from discard inference into a pixel-level comprehensive uncertainty under the approximate independence assumption using the root mean square formula. This expands the uncertainty quantification dimension from a single statistical extrapolation to three dimensions: statistics, lithological parameters, and model structure, forming a full-link uncertainty constraint. Based on this comprehensive uncertainty, the adaptive reliability distance threshold, and the percentile truncation normalized advantage score, resource target classification is performed on the target area, making the classification results more robust and reliable.

[0025] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0026] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart of a method for quantitatively assessing geothermal resource uncertainty provided in an embodiment of the present invention; Figure 2 This is a decomposition diagram of three-source uncertainty provided in an embodiment of the present invention; Figure 3 This is a percentile normalization comparison chart provided in an embodiment of the present invention; Figure 4 This is a structural diagram of a geothermal resource uncertainty quantitative assessment system provided in an embodiment of the present invention. Detailed Implementation

[0027] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] This invention provides a quantitative assessment method for geothermal resource uncertainty. Please refer to [link / reference]. Figure 1 This is a flowchart of a method for quantitatively assessing geothermal resource uncertainty provided in an embodiment of the present invention. The method includes the following steps S10-S50.

[0029] S10: Extract surface temperature anomalies in the target area from thermal infrared images.

[0030] In specific implementation of step S10, multiple scenes of the target area can be retrieved from Landsat 8 / 9 thermal infrared imagery. The DN values ​​are converted to surface temperature according to the radiometric calibration coefficient. After removing cloud and cloud image pixels using a quality mask, the pixel-level median is taken to obtain the composite temperature. Piecewise linear or robust regression fitting is performed on the composite temperature with the digital elevation model, vegetation index, and topographic factors to remove shallow thermal biases caused by elevation, vegetation, and slope aspect, resulting in a remotely sensed surface temperature anomaly. This anomaly is not interpreted as deep heat flow alone, but rather as an auxiliary thermal anomaly indicator factor in statistical regression and favorability synthesis. When soil or shallow cover thermal resistance data are available, it can be further processed according to... abnormal surface temperature Converted to surface thermal resistance Corrected equivalent heat flux indication .

[0031] It should be noted that the target area refers to the continuous geographical area for which a quantitative assessment of geothermal resource uncertainty needs to be conducted. Its spatial boundaries are pre-defined by the exploration task and can be a basin, a volcanic geothermal area, a geothermal field, or any two-dimensional rectangular or irregular polygonal area with clearly defined latitude and longitude boundaries. The target area must contain at least several measured geothermal flow observation points, layered lithospheric interface depth data, and available thermal infrared remote sensing image coverage.

[0032] S20: Using surface temperature anomaly, fault distance, and lithosphere thickness as feature matrices, and measured geothermal heat flow observation points as point source constraints, the predicted heat flow grid and its statistical regression uncertainty are obtained through regularized regression; the predicted heat flow grid is clipped to a set physical range, and pixels exceeding the boundary are marked for reliability degradation; the uncertainty is propagated according to preset rules to obtain the statistical uncertainty after physical range constraints.

[0033] In step S20, the surface temperature anomaly synthesized from multi-temporal satellite thermal infrared remote sensing images is used as an auxiliary indicator factor, and the measured geothermal heat flow observation points are used as point source constraints. The predicted heat flow grid and its statistical uncertainty are obtained through regularized regression, and the reliability of the prediction results is marked under physical range constraints.

[0034] Specifically, using the surface temperature anomaly, fault distance, and lithospheric thickness as feature matrices, Tikhonov regularized ridge regression is performed on the measured geothermal heat flow observation points. The regularization parameter is automatically selected by the generalized cross-validation criterion. Inverse distance weighted residuals are corrected by overlaying the predicted grid. Pixel-level statistical regression uncertainty is calculated using the regression parameter variance and residual spatial decay formula. This uncertainty is not forced to approach zero near the observation point, but rather maintains the lower bound of the observation error and model fitting error, asymptotically approximating the residual variance at a greater distance. The prediction results are cropped to the physical range. Then, pixels exceeding the boundary are marked for reliability degradation, and uncertainties are propagated according to conservative rules: ; in This is the lower limit of observation error. This is the penalty coefficient that increases with the distance of the boundary crossing. One possible linear form is: ; in, This is the penalty intensity coefficient, which defaults to 1 to 3. For the predicted surface heat flow, This represents the physical upper limit of surface heat flow. This represents the physical lower limit of surface heat flow. The uncertainty remains unchanged for pixels falling within this physical limit. This avoids falsely assigned narrow confidence intervals for out-of-bounds predictions due to the soft-clipping function slope being less than one.

[0035] S30: Construct a multi-layered lithospheric thermal parameter structure based on the layered interface data of the target region. Obtain the lithological parameter uncertainty by perturbing the thermal conductivity and heat generation rate of each layer and performing multi-layered heat conduction forward modeling repetitive calculations.

[0036] In step S30, a multi-layered lithospheric thermal parameter structure is constructed based on the layered interface data of the target area. The lithological parameter uncertainty is obtained through repeated calculations using parameter perturbation and multi-layered heat conduction forward modeling. Specifically, based on the layered interface depth data of the target area, a four-layered thermal parameter structure is constructed for each pixel, comprising a sedimentary layer, upper crust, lower crust, and lithospheric mantle. The thickness of each layer is obtained by subtracting the depths of adjacent interfaces, and the thermal conductivity and heat production rate of each layer are assigned values ​​from the geological lithological thermal parameter table or take default values. This forward model is based on the lumped parameter approximation solution of the one-dimensional steady-state heat conduction equation in a layered homogeneous medium, and is suitable for areas where lateral heat flow and strong groundwater convection are not dominant. In volcanic conduits, basin edges, or areas with drastic changes in heat flow gradients, the model output should be interpreted as a lower bound estimate of parameter uncertainty, and can be conservatively downgraded by combining lateral heat flow sensitivity indicators.

[0037] For example, for Layered lithosphere, each layer thickness Thermal conductivity Volumetric heat production rate At the Earth's surface heat flow Under the assumption of no lateral heat flow, the total temperature difference from the Earth's surface to the bottom of the lithosphere is contributed by two parts: the temperature drop caused by the heat flow passing through the thermal resistance of each layer. (where thermal resistance) ), and the temperature rise caused by radioactive heat generation within each layer. The average temperature rise at the center depth of each layer is used to represent the contribution of radioactive heat generation within the layer. This approximation is applicable to scenarios with relatively small layer thickness or low heat generation rate, provided that the thermal conductivity and heat generation rate are uniform within the layer (the strictly analytical solution is a parabolic temperature distribution within the layer). From this, the relationship between the top-to-bottom temperature difference and the surface heat flow is obtained: in, As surface heat flow, For the index of the lithosphere layers, This represents the equivalent top-to-bottom temperature difference in the lithosphere.

[0038] The equivalent top-to-bottom temperature difference of the lithosphere is inferred from the current predicted heat flow and multilayer thermal parameters. This equivalent top-to-bottom temperature difference is used as a conditional sensitivity constraint based on the current best estimate, rather than a strictly fixed bottom boundary physical quantity. Subsequently, Gaussian relative perturbations are applied to the thermal conductivity and heat production rate of each layer, and the perturbed surface heat flow is repeatedly calculated under this conditional sensitivity framework using the multilayer thermal resistance forward modeling formula. in, For the perturbation of the first Layer thermal conductivity, For the perturbation of the first Layer volumetric heat generation rate.

[0039] The pixel-by-pixel standard deviation of multiple realizations is taken as the uncertainty of lithological parameters. When the layered interface data is unavailable, it degenerates into a single-layer equivalent model.

[0040] S40: Construct a physical information neural network with spatial coordinates as input and temperature as output. Each hidden layer of the physical information neural network contains a dropout layer. During training, the loss function incorporates the following constraints: remotely sensed surface temperature Dirichlet constraint corrected for shallow bias, measured heat flux Neumann flux constraint, well temperature internal point constraint, variable-depth lithosphere bottom interface temperature constraint, layered steady-state heat conduction physical residual term, and soft constraint that temperature monotonically increases with depth. After training, the dropout layer is kept active and multiple forward propagations are performed. The model uncertainty is obtained based on the pixel-by-pixel standard deviation of the multiple forward propagations.

[0041] In step S40, a stochastic reasoning mechanism is introduced into the physical information neural network to obtain the model uncertainty through multiple forward propagations.

[0042] Specifically, a multilayer perceptron is constructed with spatial coordinates (longitude, latitude, and depth) as input and temperature as output to fit a three-dimensional steady-state temperature field. A dropout layer is set after each hidden layer, and the dropout rate is a configurable parameter with a typical range of 3% to 15%. For physical information networks that need to maintain the satisfaction of physical constraints, a lower value such as 5% is recommended, while a higher value can be used for purely data-driven networks.

[0043] The loss function incorporates the following terms: remotely sensed surface temperature Dirichlet constraint corrected for shallow bias, measured heat flux Neumann flux constraint, borehole temperature internal point constraint, variable-depth lithosphere bottom interface temperature constraint, stratified steady-state heat conduction physical residual term, and soft constraint that temperature monotonically increases with depth. Among these: The bottom interface depth function used for the temperature constraint of the variable-depth lithosphere bottom interface is obtained by spatial interpolation from the lithosphere thickness grid: regular grids use bilinear interpolation, while irregular grids automatically degenerate into inverse distance weighted or triangular mesh linear interpolation. The soft constraint, where temperature monotonically increases with depth, is weighted by an adjustable coefficient. Add to the total loss, and only apply to points below the preset shallow depth.

[0044] During inference, the discard layer remains active, and forward propagation is performed a preset number of times (typically 30-100 times). The pixel-by-pixel standard deviation of the forward propagation results is used as the model uncertainty. .

[0045] S50: Combine statistical uncertainty, lithological parameter uncertainty and model uncertainty into a pixel-level comprehensive uncertainty, and classify the target area into resource targets based on the comprehensive uncertainty, adaptive reliability distance threshold and percentile truncation normalized goodness score.

[0046] In this embodiment, the three sources of uncertainty are combined according to the root mean square under the approximate independence assumption: in, To account for the overall uncertainty, The statistical regression uncertainty is constrained by the physical range. For the uncertainty of lithological parameters, This represents the model uncertainty.

[0047] The above merging method is used to give an approximate comprehensive uncertainty. When the covariance between the three sources of error can be estimated, it can be expanded into a quadratic merging that includes the covariance matrix. When the covariance cannot be estimated, it is treated as an engineering approximation and identified in the result metadata.

[0048] In some embodiments, near-end, mid-end, and far-end reliability distance thresholds are adaptively calculated based on the median nearest neighbor distance of measured heat flow observation points, allowing the thresholds to be dynamically adjusted with observation density: the near-end radius is narrowed in densely populated areas and moderately widened in sparsely populated areas while retaining a higher uncertainty indicator. Each scoring index is normalized using a percentile truncation method to suppress the compression effect of extreme outliers on the scoring discrimination power of normal areas. The combined uncertainty, adaptive reliability distance thresholds, and percentile truncation normalization jointly participate in the resource target classification and grading constraints.

[0049] Specifically, the target area is classified into resource targets through the following steps S501-S503.

[0050] S501: Dynamically calculate the near-end, middle-end, and far-end reliability distance thresholds based on the median nearest neighbor distance of the measured heat flow observation points as adaptive reliability distance thresholds.

[0051] For all measured geothermal flow observation points within the target area, the spatial distance between each observation point and its nearest neighbor is calculated, resulting in a set of nearest neighbor distance values. The median of this set of distance values ​​is taken as the adaptive distance benchmark. The near-end reliability distance threshold is set to 0.2–0.4 times this benchmark, the mid-end reliability distance threshold is set to 0.8–1.2 times this benchmark, and the far-end reliability distance threshold is set to 2–3 times this benchmark. When the observation points are extremely dense, causing the near-end threshold to be less than the grid resolution or the minimum geological unit size, the near-end threshold can be set to that resolution or the minimum geological unit size as the lower limit. For any pixel within the target area, its spatial distance to the nearest observation point is calculated, and the reliability level of the pixel is determined based on which of the above threshold intervals the distance falls into.

[0052] S502: Use preset low percentile and high percentile thresholds to truncate and normalize each item of the favorability score as percentile truncation normalization.

[0053] For each sub-item in the favorableness score, the numerical distribution of all pixels within the target area for that sub-item is statistically analyzed. A lower percentile is selected as the lower truncation limit, which can be chosen between the 1st and 5th percentiles; a higher percentile is selected as the upper truncation limit, which can be chosen between the 95th and 99th percentiles. Values ​​in the original pixel values ​​below the lower truncation limit are forcibly set to the lower truncation limit value, and values ​​above the upper truncation limit are forcibly set to the upper truncation limit value. Then, the truncated values ​​are linearly scaled using a minimum-maximum scaling method to map them to a normalized interval of zero to one. In this way, the numerical differences of most normal pixels distributed between the two truncation percentiles can be fully preserved, while extreme outliers far from the normal distribution range are truncated, without compressing the normalized score interval of normal pixels.

[0054] S503: Based on the comprehensive uncertainty, the reliability distance threshold level of each pixel's distance to the nearest observation point, and the favorable score after percentile truncation and normalization, the target area is divided into multiple resource levels, among which pixels with comprehensive uncertainty exceeding the preset threshold are downgraded.

[0055] For each pixel within the target area, its comprehensive uncertainty, the reliability threshold level corresponding to its distance to the nearest measured heat flow observation point, and the favorability score obtained after percentile truncation and normalization are obtained. Multiple resource levels are set, each corresponding to a threshold range for the favorability score, a minimum requirement for the reliability level, and an upper limit requirement for the comprehensive uncertainty. The threshold range for the favorability score can be determined using the equal-interval or natural breakpoint method; the minimum requirement for the reliability level can be set according to the risk tolerance of the exploration stage, for example, it can be relaxed in the preliminary exploration stage and stricter in the detailed exploration stage; the upper limit requirement for the comprehensive uncertainty can be determined based on the average comprehensive uncertainty at the observation point, the standard deviation of the regional background heat flow, or the exploration investment risk threshold. Furthermore, an absolute downgrade threshold for comprehensive uncertainty is set, which can be set to 1.5 to 2 times the mean of the comprehensive uncertainty of all observation points in the region, or set to the maximum uncertainty tolerance allowed for exploration decisions. If the comprehensive uncertainty of any pixel exceeds this absolute downgrade threshold, regardless of its favorability score and reliability level, it will be forcibly downgraded.

[0056] To further illustrate the feasibility and progressiveness of the method of the present invention, two examples are provided below for detailed explanation.

[0057] Example 1: Regional end-to-end uncertainty assessment under complete hierarchical interface data.

[0058] The target area is a volcanic geothermal region, covering an area of ​​approximately 15,000 km². 2The remote sensing data source consisted of six Landsat 9 thermal infrared images from June to September 2023. These images were composited using cloud masking and median data to obtain a surface temperature raster with an effective pixel rate of 93%. The digital elevation model (DEM) showed the target area's elevation range to be 1600–4200 m. Surface temperature anomalies were obtained by piecewise linear fitting combined with vegetation indices to remove differences in elevation and shallow cover. These anomalies served as thermal anomaly indicators in subsequent regressions and were not interpreted as pure deep heat flow on their own.

[0059] The measured data includes data from 38 geothermal flow stations (range 45~125 mW / m). 2 The data includes depth data of the target area's stratification interfaces (bottom interface of sedimentary layers 0.5–4 km, bottom interface of upper crust 12–20 km, Moho discontinuity 28–38 km, bottom interface of the lithosphere (LAB) 65–110 km). The geological lithology and thermal parameter table covers four lithologies: granite, limestone, sandstone, and volcanic rocks.

[0060] The ridge regression regularization parameter was selected as 0.15 after generalized cross-validation, and the predicted heat flux range was 35–122 mW / m. 2 Among them, it is 45mW / m lower than the measured minimum value. 2 The region located in the extrapolation zone far from the station is assigned higher uncertainty and marked as low reliability in the reliability grading. Statistical regression uncertainty The spatial average is 7.2 mmW / m. 2 .

[0061] The system constructs a four-layer thermal parameter structure for each pixel: the sedimentary layer thickness is taken from the depth of the sedimentary layer's bottom interface (where the land water depth is zero); the upper crust thickness equals the upper crust bottom interface depth minus the sedimentary layer's bottom interface depth; the lower crust thickness equals the Moho depth minus the upper crust bottom interface depth; and the lithospheric mantle thickness equals the lithospheric bottom interface depth minus the Moho depth. The thermal conductivity of each layer is assigned spatially from the geological map, overriding the default values. A relative standard deviation of 15% for thermal conductivity and 25% for heat production rate are set, and 200 Gaussian perturbation forward models are performed. The results show that the lithological parameter uncertainty... The average is 9.8 mW / m 2 Spatial analysis shows: high heat flux region (greater than 100 mW / m) 2 (average) It is 14.2mW / m 2 Low heat flux region (less than 70mW / m) 2 The value is 6.1 mW / m 2 This reflects the nonlinear thermal resistance effect, making the high-temperature region more sensitive to parameter deviations. Changes in the deposition layer thickness will... The identifiable differences, in their specific direction and magnitude, depend on the thermal conductivity of the deposited layer, the thickness distribution of adjacent layers, and the temperature difference between the top and bottom of the conditions. The mechanism is that the low thermal conductivity deposited layer changes the interlayer weight in the total thermal resistance.

[0062] The constructed physical information neural network consists of four hidden layers, each with 64 units, followed by a dropout layer with a dropout rate of 5%. Before training, sensitivity comparisons were performed for dropout rates of 0%, 3%, 5%, 10%, and 15%. A 5% dropout rate resulted in lower validation error and more stable physical residual convergence, and was therefore chosen for this embodiment. After 1500 epochs of training, the total loss converged to 48. During inference, dropout activations were maintained, and 50 forward propagations were performed.

[0063] Discarding inference uncertainty The concentration at the center of the target area (dense observation area) is 2.3 to 4.1 mW / m. 2 The strength increases to 9 to 16 mW / m at the edge of the target area (far from the extrapolated area of ​​the observation). 2 This aligns with Bayesian inference theory.

[0064] After combining statistical uncertainty, lithological parameter uncertainty, and discard inference uncertainty using the root mean square method, the overall uncertainty is... The spatial average is 13.5 mW / m 2 The maximum value of 28.7 mW / m was observed in areas far from all observation points. 2 .

[0065] The adaptive distance threshold was calculated based on the median nearest neighbor distance of 18.5 km for 38 observation points, resulting in three levels of thresholds: 5.0 km for the near end, 20.0 km for the middle end, and 50.0 km for the far end. The 5 km near-end threshold was set to prevent threshold degradation to the grid resolution level in extremely dense observation scenarios. Compared to fixed thresholds (15, 50, 100 km), the adaptive near-end, middle-end, and far-end thresholds in this embodiment are tightened by approximately 67%, 60%, and 50%, respectively. Reliability grading results: scores greater than 0.8 for distances within 5 km of the observation point, 0.4–0.8 for distances between 5 and 20 km, and less than 0.2 for distances exceeding 50 km.

[0066] In the resource classification constraints, each component of the favorable score is normalized to percentiles (truncated at the 2nd and 98th percentiles). Extreme high-temperature pixels in the crater do not affect the score of normal areas. Pixels with uncertainty exceeding 25 milliwatts per square meter are downgraded, accounting for 8.3% of the target area at level 2 and above, all located in the far-end extrapolation zone.

[0067] In a comparative analysis, when the surface temperature anomaly component of the same target area was normalized using min-max normalization, the crater pixel (with a surface temperature anomaly 15 times the background mean) compressed the remaining 98% of pixels into the 0-0.07 range, resulting in a loss of discriminative power in the favorability score. After applying percentile normalization (truncating the 2nd and 98th percentiles), the effective pixel scores were evenly distributed in the 0-1 range, and the interquartile range improved by approximately one order of magnitude, clearly distinguishing between moderate and weak anomaly areas. This comparison verifies the necessity of percentile normalization in scenarios containing extreme outliers.

[0068] Example 2: Degeneracy scenario with sparse observations and missing hierarchical interfaces.

[0069] In a preliminary reconnaissance area, only six heat flow stations were available, and no stratification interface data was found. The system detected that the stratification interface was unavailable and automatically switched to a single-layer equivalent forward modeling mode: perturbation was performed using the weighted average thermal conductivity of the lithological distribution on the geological map and the effective conductive layer thickness obtained from regional sedimentary thickness statistics; when sedimentary thickness data was also unavailable, a default equivalent thickness of 5 km and an average thermal conductivity of 2.5 W / (m·K) from the geological map were used. In this embodiment, due to the lack of sedimentary thickness data, default parameters were used for perturbation. The average is 11.1 mW / m 2 .

[0070] The median nearest neighbor distance for the six observation points is 30.9 km. Adaptive thresholds: 7.7 km for the near end, 30.9 km for the middle end, and 77.3 km for the far end. Compared with fixed thresholds (15, 50, 100 km): the adaptive results are more stringent at the near end and more lenient at the far end, with a more continuous and smooth transition in reliability.

[0071] Due to the limited number of observation points, the overall inference uncertainty of physical information neural networks is relatively large (average 12.5 mW / m). 2 This accurately reflects the cognitive uncertainty of the model under sparse data. The average value after end-to-end merging... It is 18.1mW / m 2 This is higher than Example 1, reflecting the limited reliability under sparse observations.

[0072] Example 2 demonstrates that each module of the system can degrade when data is limited: single-layer forward modeling provides conservative but effective parameter uncertainty, adaptive distance adapts to sparse patterns, and percentile normalization is equally robust when the amount of data is small.

[0073] like Figure 2As shown, the statistical uncertainty, lithological parameter uncertainty, and model uncertainty after physical range constraints in Examples 1 and 2 can all be combined into a comprehensive uncertainty using the root mean square method. The model uncertainty and comprehensive uncertainty are higher under sparse observation conditions, indicating that the present invention can reflect the uncertainty differences under different data constraints.

[0074] like Figure 3 As shown, in the presence of pixels with extreme surface temperature anomalies, min-maximum normalization compresses the scoring range of normal pixels, while percentile truncation normalization preserves the scoring differences within the normal region, thereby improving the spatial discrimination of resource target classification.

[0075] The key values ​​in the two examples above were verified by arithmetic.

[0076] Example 1: The median nearest neighbor distance for 38 observation points is 18.5 km, and the adaptive threshold can be reproduced at 5.0, 20.0, and 50.0 km. The comprehensive uncertainty is calculated by combining the statistical uncertainty (7.2), lithological parameter uncertainty (9.8), and discarded inference uncertainty (approximately 3.2 in the central region and approximately 12.5 in the peripheral region) at the root mean square, with a spatial average of approximately 13.5 mW / m. 2 The extrapolation range can exceed 25mW / m 2 The downgrade threshold.

[0077] Example 2: The median nearest neighbor distance for 6 observation points is 30.9 km, and the adaptive threshold can reproduce distances of approximately 7.7, 30.9, and 77.3 km; the single-layer equivalent Monte Carlo forward modeling uncertainty is 11.1 mW / m 2 Being in the same order of magnitude, the combined three sources can reproduce 18.1 mW / m. 2 The overall uncertainty is relatively high. Verification results show that the key numerical relationships in the two embodiments are consistent with the algorithm formula.

[0078] Another aspect of the present invention provides a quantitative assessment system for geothermal resource uncertainty, such as... Figure 4 The diagram shown is a structural diagram of a geothermal resource uncertainty quantitative assessment system provided in an embodiment of the present invention. The geothermal resource uncertainty quantitative assessment system includes: Data extraction module 401 is configured to extract surface temperature anomalies of the target area from thermal infrared images; The first uncertainty calculation module 402 is configured to use the surface temperature anomaly, fault distance, and lithosphere thickness as feature matrices, and measured geothermal heat flow observation points as point source constraints, to obtain the predicted heat flow grid and its statistical regression uncertainty through regularized regression; to clip the predicted heat flow grid to a set physical range, and to mark the reliability degradation of pixels that exceed the boundary; to propagate the uncertainty according to preset rules, and to obtain the statistical uncertainty after physical range constraints. The second uncertainty calculation module 403 is configured to construct a multi-layer lithospheric thermal parameter structure based on the layered interface data of the target region, and obtain the lithological parameter uncertainty by applying perturbation to the thermal conductivity and heat generation rate of each layer and performing multi-layer heat conduction forward modeling repeated calculations. The third uncertainty calculation module 404 is configured to construct a physical information neural network with spatial coordinates as input and temperature as output. Each hidden layer of the physical information neural network contains a dropout layer. During training, the loss function incorporates the following constraints: remotely sensed surface temperature Dirichlet constraint corrected for shallow bias, measured heat flux Neumann flux constraint, drilling temperature internal point constraint, variable-depth lithosphere bottom interface temperature constraint, layered steady-state heat conduction physical residual term, and soft constraint that temperature monotonically increases with depth. After training, the dropout layer is kept active and multiple forward propagations are performed. The model uncertainty is obtained based on the pixel-by-pixel standard deviation of the multiple forward propagations. The target classification module 405 is configured to merge the statistical uncertainty, lithological parameter uncertainty and model uncertainty into a pixel-level comprehensive uncertainty, and classify the target area based on the comprehensive uncertainty, the adaptive reliability distance threshold and the percentile truncation normalized advantage score.

[0079] Another aspect of the present invention provides an electronic device, comprising: a controller; and a memory for storing one or more programs, which, when executed by the controller, perform the methods described in the various embodiments above.

[0080] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described above. This computer-readable storage medium may be included in the electronic device described in the above embodiments, or it may exist independently without being assembled into the electronic device.

[0081] Another aspect of the present invention provides a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various embodiments described above.

[0082] According to one aspect of the present invention, a computer system is also provided, including a central processing unit (CPU) that can perform various appropriate actions and processes, such as executing the methods described above, based on a program stored in read-only memory (ROM) or a program loaded from storage into random access memory (RAM). The RAM also stores various programs and data required for system operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0083] The following components are connected to the I / O interface: input components including keyboards, mice, etc.; output components including cathode ray tubes (CRTs), liquid crystal displays (LCDs), and speakers; storage components including hard drives; and communication components including network interface cards such as LAN (Local Area Network) cards and modems. The communication components perform communication processing via networks such as the Internet. Drives are also connected to the I / O interface as needed. Removable media, such as disks, optical discs, magneto-optical discs, semiconductor memories, etc., are installed on the drive as needed so that computer programs read from them can be installed into the storage components as required.

[0084] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs various functions defined in the system of the present invention.

[0085] It should be noted that the computer-readable medium shown in the embodiments of the present invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the present invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable computer program. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The computer program contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0086] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0087] The module units described in the embodiments of the present invention can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0088] The above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the present invention, and the patent protection scope of the present invention should be defined by the claims.

Claims

1. A quantitative assessment method for geothermal resource uncertainty, characterized in that, The method includes: Extracting surface temperature anomalies in the target area from thermal infrared images; Using the surface temperature anomaly, fault distance, and lithosphere thickness as feature matrices, and measured geothermal heat flow observation points as point source constraints, a predicted heat flow grid and its statistical regression uncertainty are obtained through regularized regression. The predicted heat flow grid is then cropped to a set physical range, and pixels exceeding the boundary are marked with reliability degradation. Uncertainty is propagated according to preset rules to obtain the statistical uncertainty after physical range constraints. Based on the layered interface data of the target region, a multi-layered lithosphere thermal parameter structure is constructed. By applying perturbation to the thermal conductivity and heat generation rate of each layer and performing multi-layered heat conduction forward modeling and repeated calculations, the uncertainty of lithological parameters is obtained. A physical information neural network is constructed with spatial coordinates as input and temperature as output. Each hidden layer of the physical information neural network contains a dropout layer. During training, the loss function incorporates the following constraints: remotely sensed surface temperature Dirichlet constraint corrected for shallow bias, measured heat flux Neumann flux constraint, borehole temperature internal point constraint, variable-depth lithosphere bottom interface temperature constraint, layered steady-state heat conduction physical residual term, and soft constraint that temperature monotonically increases with depth. After training, the dropout layer is kept active and multiple forward propagations are performed. The model uncertainty is obtained based on the pixel-by-pixel standard deviation of the multiple forward propagations. The statistical uncertainty, lithological parameter uncertainty, and model uncertainty are combined into a pixel-level comprehensive uncertainty. Based on the comprehensive uncertainty, the adaptive reliability distance threshold, and the percentile truncation normalized advantage score, resource target classification is performed on the target area.

2. The quantitative assessment method for geothermal resource uncertainty according to claim 1, characterized in that, The statistical uncertainty after physical range constraints is calculated using the following formula. : in, This is the lower limit of observation error. This is the penalty coefficient that increases with the distance of the boundary crossing. For pixel-level statistical regression uncertainty, It is a function for maximizing the value; The penalty coefficient Represented as: in, The penalty intensity coefficient, The predicted surface heat flow value, This represents the physical upper limit of surface heat flow. This represents the physical lower limit of surface heat flow.

3. The method for quantitative assessment of geothermal resource uncertainty according to claim 1, characterized in that, Based on the layered interface data of the target region, a multi-layered lithospheric thermal parameter structure is constructed. By perturbing the thermal conductivity and heat production rate of each layer and performing multi-layered heat conduction forward modeling and repeated calculations, the lithological parameter uncertainties are obtained, including: for Layered lithosphere, obtaining the thickness of each layer Thermal conductivity and volumetric heat production rate Based on the current predicted heat flow and multilayer thermal parameters, the equivalent top-to-bottom temperature difference of the lithosphere can be inversely calculated using the following formula. : in, As surface heat flow, The index represents the number of layers in the lithosphere; the above formula is an approximation of the equivalent thermal resistance under one-dimensional steady-state heat conduction conditions, used for parameter perturbation sensitivity estimation. After applying perturbations to the thermal conductivity and heat generation rate of each layer, the equivalent top-bottom temperature difference... Under the constraints, the disturbed surface heat flow is calculated using the following formula. : in, For the perturbation of the first Layer thermal conductivity, For the perturbation of the first Layer volumetric heat production rate; Take the result from multiple calculations The pixel-by-pixel standard deviation as the uncertainty of lithological parameters .

4. The method for quantitative assessment of geothermal resource uncertainty according to claim 1, characterized in that, The formula for determining the comprehensive uncertainty is: in, To account for the overall uncertainty, The statistical uncertainty is the result of physical range constraints. For the uncertainty of lithological parameters, This represents the model uncertainty.

5. The method for quantitative assessment of geothermal resource uncertainty according to claim 1, characterized in that, In the process of extracting surface temperature anomalies of a target area from thermal infrared images, when soil or shallow cover thermal resistance data are available, the surface temperature anomalies are converted into equivalent heat flux indicators after surface thermal resistance correction using the following formula: in, This is an equivalent heat flux indicator. This is due to abnormal surface temperature. This represents the surface thermal resistance.

6. The quantitative assessment method for geothermal resource uncertainty according to claim 1, characterized in that, The shallow bias-corrected remote sensing surface temperature surface Dirichlet constraint uses the extracted surface temperature anomaly as the upper boundary condition. The bottom interface depth function used for the temperature constraint of the variable-depth lithosphere bottom interface is obtained by spatial interpolation from the lithosphere thickness grid. Bilinear interpolation is used for regular grids, while it degenerates into inverse distance weighted or triangular mesh linear interpolation for irregular meshes. The soft constraint, in which temperature monotonically increases with depth, is added to the total loss with an adjustable weighting coefficient and only applies to collocation points below a preset shallow depth.

7. The method for quantitative assessment of geothermal resource uncertainty according to claim 1, characterized in that, Based on the comprehensive uncertainty, adaptive reliability distance threshold, and percentile truncation normalized advantage score, resource target classification and grading of the target area are performed, including: The adaptive reliability distance thresholds are dynamically calculated based on the median nearest neighbor distance of the measured heat flow observation points, and are used as the near-end, middle-end, and far-end reliability distance thresholds. The percentile truncation normalization is obtained by truncating and normalizing each item of the advantage score using preset low percentile thresholds and high percentile thresholds. Based on the comprehensive uncertainty, the reliability distance threshold level to which the distance from each pixel to the nearest observation point belongs, and the favorable score after percentile truncation and normalization, the target area is divided into multiple resource levels, among which pixels with comprehensive uncertainty exceeding the preset threshold are downgraded.

8. A quantitative assessment system for geothermal resource uncertainty, characterized in that, The system includes: The data extraction module is configured to extract surface temperature anomalies of the target area from thermal infrared images; The first uncertainty calculation module is configured to use the surface temperature anomaly, fault distance, and lithosphere thickness as feature matrices, and measured geothermal heat flow observation points as point source constraints. It obtains the predicted heat flow grid and its statistical regression uncertainty through regularized regression. The predicted heat flow grid is then clipped to a set physical range, and pixels exceeding the boundary are marked with reliability degradation. Uncertainty is propagated according to preset rules to obtain the statistical uncertainty after physical range constraints. The second uncertainty calculation module is configured to construct a multi-layered lithosphere thermal parameter structure based on the layered interface data of the target region. By applying perturbation to the thermal conductivity and heat generation rate of each layer and performing multi-layered heat conduction forward modeling repeated calculations, the uncertainty of lithological parameters is obtained. The third uncertainty calculation module is configured to construct a physical information neural network with spatial coordinates as input and temperature as output. Each hidden layer of the physical information neural network contains a dropout layer. During training, the loss function incorporates the following constraints: remotely sensed surface temperature Dirichlet constraint corrected for shallow bias, measured heat flux Neumann flux constraint, well temperature internal point constraint, variable-depth lithosphere bottom interface temperature constraint, layered steady-state heat conduction physical residual term, and soft constraint that temperature monotonically increases with depth. After training, the dropout layer is kept active and multiple forward propagations are performed. The model uncertainty is obtained based on the pixel-by-pixel standard deviation of the multiple forward propagations. The target classification module is configured to merge the statistical uncertainty, lithological parameter uncertainty and model uncertainty into a pixel-level comprehensive uncertainty, and classify the target area based on the comprehensive uncertainty, the adaptive reliability distance threshold and the percentile truncation normalized advantage score.

9. An electronic device, characterized in that, The electronic device includes: Memory, used to store computer programs; A processor for executing the computer program to implement the method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium storing instructions, characterized in that, When the instructions are executed by the processor, the method according to any one of claims 1 to 7 is performed.

Images