Method and device for constructing coupling model for mine ecological restoration effect prediction

Abstract

This invention provides a method and apparatus for constructing a coupled model for predicting the effects of mine ecological restoration, relating to the field of artificial intelligence technology. It divides multi-source heterogeneous information on mine ecological restoration into dynamic temporal features and static spatial features (topography, soil, hydrology, etc.). Based on the spatial constraints of the static features, it constrains and integrates the local variation patterns of the dynamic temporal features across multiple time scales, thereby uncovering the spatiotemporal coupling mechanism between the two. This allows for the construction of a multi-factor coupled climate and environmental impact feature reflecting the combined effects of engineering measures, site conditions, and climate fluctuations. Further model construction enables the neural network model to fully learn the nonlinear mapping relationship and long-term evolution patterns between influencing factors and restoration effects, forming a customized prediction model for the target mine. This allows for the deduction and quantitative prediction of the dynamic evolution trend of future ecological restoration effects, improving the scientific rigor, stability, and timeliness of the prediction.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for constructing a coupled model for predicting the effects of mine ecological restoration. Background Technology

[0002] Mineral resources, as a crucial material foundation for industrial production and socio-economic development, have greatly promoted social progress through their continuous development and utilization. However, mining activities have also triggered a series of serious ecological and environmental problems, among which land damage, vegetation destruction, and soil erosion are the most typical and prominent. To alleviate the ecological damage caused by mining and achieve coordinated development between the ecological environment and mineral development, various mine ecological restoration projects have been widely carried out both domestically and internationally in recent years. These projects employ a series of engineering measures, such as vegetation reconstruction, soil improvement, and landform reshaping, to attempt to repair damaged ecosystems and restore their ecological service functions. However, the formation and evolution of mine ecological restoration effects is a long-term, dynamic, and complex process. The effectiveness of restoration is influenced by multiple factors, primarily including the technical measures adopted in the restoration project, the site conditions of the mine itself (such as soil texture, topography, and hydrological conditions), and regional climate fluctuations (such as precipitation and temperature changes). Due to the uncertainty and interrelationship of these factors, the evolution of restoration effects over time exhibits significant complexity and unpredictability.

[0003] Currently, existing mine ecological restoration work largely focuses on the implementation process of restoration projects, with assessments of restoration effects concentrated on short-term static monitoring. There is a lack of accurate prediction of the dynamic evolution of restoration effects over the next few years or even decades. Accurate prediction of the long-term dynamic evolution of mine ecological restoration effects is a crucial prerequisite for scientifically evaluating the long-term effectiveness of restoration plans, optimizing project resource allocation, avoiding resource waste, and achieving precise and intelligent restoration management. It can provide important decision support for subsequent optimization and long-term management of mine ecological restoration projects, and is of paramount importance for improving the scientific rigor and effectiveness of mine ecological restoration. Therefore, given the lack of existing technologies for long-term dynamic prediction of mine ecological restoration effects, there is an urgent need to develop a technical solution capable of accurately predicting the evolution trend of restoration effects to address the aforementioned technical problems. Summary of the Invention

[0004] The purpose of this invention is to provide a method and apparatus for constructing a coupled model for predicting the effect of mine ecological restoration, which can realize the dynamic evolution trend deduction and quantitative prediction of the effect of mine ecological restoration.

[0005] In a first aspect, embodiments of the present invention provide a method for constructing a coupled model for predicting the effects of mine ecological restoration. The method includes: acquiring engineering implementation parameters of a mine ecological restoration project, as well as site condition parameters and continuous climate and environmental monitoring sequences corresponding to the engineering implementation parameters; acquiring dynamic temporal features and static features from the engineering implementation parameters, site condition parameters, and continuous climate and environmental monitoring sequences; constructing coupled features of the climate and environmental impact of the mine ecological restoration project based on the spatial constraint information of the static features and the temporal local features of the dynamic temporal features at multiple time scales; constructing a training sample set for an initial neural network based on the coupled features of the climate and environmental impact; and training the initial neural network using the training sample set to construct a prediction model for the ecological restoration effects of the mine ecological restoration project.

[0006] In conjunction with the first aspect, this invention also provides a first implementation of the first aspect, wherein the step of constructing climate and environmental impact coupling features for mine ecological restoration projects based on spatial constraint information of static features and temporal local features of dynamic temporal features at multiple time scales includes: using a multilayer perceptron to perform feature mapping on the spatial constraint information of static features to obtain a context vector characterizing the ecological response characteristics of static features; based on the context vector, determining the scale attention weights corresponding to the temporal local features of dynamic temporal features at each time scale; the scale attention weights are used to characterize the degree of influence of the temporal local features at the corresponding time scale on the ecological restoration effect; based on the scale attention weights, performing dynamic weighted fusion processing on the temporal local features at all time scales to obtain a dynamic climate fusion feature vector; and performing feature coupling processing on the static features and the dynamic climate fusion feature vector to construct the climate and environmental impact coupling features for mine ecological restoration projects.

[0007] In conjunction with the first aspect, this embodiment of the invention also provides a second implementation of the first aspect, wherein the step of determining the scale attention weights corresponding to the temporal local features of the dynamic temporal features at each time scale based on the context vector includes: determining the scale attention weights of the temporal local features at the corresponding time scale based on the dot product similarity between the temporal local features and the context vector at each time scale.

[0008] In conjunction with the first aspect, this embodiment of the invention also provides a third implementation of the first aspect, wherein the step of performing feature coupling processing on static features and dynamic climate fusion feature vectors to construct climate and environmental impact coupling features of mine ecological restoration projects includes: performing bilinear interactive coupling processing on static features and dynamic climate fusion feature vectors based on a bilinear interactive attention mechanism to construct climate and environmental impact coupling features of mine ecological restoration projects.

[0009] In conjunction with the first aspect, this invention also provides a fourth implementation of the first aspect, wherein the steps of constructing a training sample set for an initial neural network based on the coupling characteristics of climate and environmental impacts, and using the training sample set to train the initial neural network to construct a prediction model for the ecological restoration effect corresponding to the mine ecological restoration project, include: constructing a training sample set corresponding to the coupling characteristics of climate and environmental impacts based on the restoration effect indicators corresponding to the coupling characteristics of climate and environmental impacts; determining the restoration effect memory state vector corresponding to the coupling characteristics of climate and environmental impacts using a time-varying decay memory gate; training the initial neural network based on the restoration effect memory state vector, the training sample set, and preset engineering and climate matching degree features; and adjusting the loss weights of the initial neural network based on the prediction uncertainty of the model training to construct a prediction model for the ecological restoration effect corresponding to the mine ecological restoration project.

[0010] In conjunction with the first aspect, this invention also provides a fifth implementation of the first aspect, wherein the initial neural network is trained based on a loss function, the loss function is constructed based on the Charbonnier loss function and temporal decay weights, the loss function includes multiple repair effect prediction targets, and the loss weight of each repair effect prediction target is calculated based on the corresponding prediction uncertainty.

[0011] In conjunction with the first aspect, this invention also provides a sixth implementation method of the first aspect, wherein the method further includes: dividing the engineering implementation parameters, site condition parameters, and climate environment continuous monitoring sequences into multiple groups of data to be processed based on the data characteristics of the engineering implementation parameters, site condition parameters, and climate environment continuous monitoring sequences; performing data normalization processing on each group of data to be processed to construct a standardized feature matrix corresponding to the mine ecological restoration project.

[0012] In conjunction with the first aspect, this embodiment of the invention also provides a seventh implementation of the first aspect, wherein the data characteristics include: continuous numerical characteristics, classification category characteristics, and monitoring time series characteristics; multiple sets of data to be processed include continuous parameters corresponding to continuous numerical characteristics, discrete classification parameters corresponding to classification category characteristics, and climate monitoring data corresponding to monitoring time series characteristics; the step of performing data normalization processing on each set of data to be processed to construct a standardized feature matrix corresponding to the mine ecological restoration project includes: based on the data distribution of continuous parameters and the degree of physical influence of continuous parameters on the ecological restoration effect, performing adaptive physical constraint normalization processing on the corresponding continuous parameters to generate first preprocessed data; based on the one-hot encoding vector corresponding to the discrete classification parameters, performing norm normalization processing on the corresponding discrete classification parameters to generate second preprocessed data; based on the continuous feature statistical parameters of the climate monitoring data, performing adaptive physical constraint normalization processing on the climate monitoring data to generate third preprocessed data; and concatenating the first preprocessed data, the second preprocessed data, and the third preprocessed data to construct a standardized feature matrix corresponding to the mine ecological restoration project.

[0013] Secondly, embodiments of the present invention provide a method for dynamically predicting the ecological restoration effect of a mine. The method includes: acquiring ecological restoration project data of a mine to be predicted; inputting the ecological restoration project data of a mine to a pre-constructed ecological restoration effect prediction model; and outputting the dynamic prediction result of the ecological restoration effect of the mine ecological restoration project data. The ecological restoration effect prediction model is constructed based on any of the above embodiments. Based on the dynamic prediction result of the restoration effect, an ecological restoration project scheme for the target mine is determined.

[0014] Thirdly, embodiments of the present invention provide a coupled model construction device for predicting the effect of mine ecological restoration. The device includes: a data acquisition module for acquiring engineering implementation parameters of the mine ecological restoration project, as well as site condition parameters and continuous climate and environmental monitoring sequences corresponding to the engineering implementation parameters; a data processing module for acquiring dynamic temporal features and static features from the engineering implementation parameters, site condition parameters, and continuous climate and environmental monitoring sequences, and constructing coupled features of the climate and environmental impact of the mine ecological restoration project based on the spatial constraint information of the static features and the temporal local features of the dynamic temporal features at multiple time scales; and an execution module for constructing a training sample set for an initial neural network based on the coupled features of the climate and environmental impact, and training the initial neural network using the training sample set to construct an ecological restoration effect prediction model corresponding to the mine ecological restoration project.

[0015] The embodiments of this invention bring the following beneficial effects: This invention provides a method and apparatus for constructing a coupled model for predicting the effects of mine ecological restoration. By dividing the multi-source heterogeneous information of mine ecological restoration projects into dynamic temporal features that evolve over time and relatively stable static spatial features such as topography, soil, and hydrology, and based on the inherent spatial constraints of the static features, the local variation patterns of the dynamic temporal features at multiple time scales are constrained and integrated. This allows for in-depth exploration of the spatiotemporal coupling mechanism between dynamic and static features, constructing a multi-factor coupled climate and environmental impact feature that comprehensively reflects the combined effects of engineering measures, site conditions, and climate fluctuations. Based on this feature, a training sample set is constructed, enabling the neural network model to fully learn the nonlinear mapping relationship and long-term evolution law between various influencing factors and restoration effects, forming a dedicated ecological restoration effect prediction model adaptable to the target mine scenario. This allows for the extrapolation and quantitative prediction of the dynamic evolution trend of ecological restoration effects over the next few years or even decades, improving the scientific rigor, stability, and timeliness of the prediction results.

[0016] Meanwhile, based on the long-term dynamic prediction results output by this model, it can also provide reliable data support and decision-making basis for the long-term effectiveness evaluation of restoration schemes, the rational allocation and dynamic regulation of engineering resources, promote the transformation of mine ecological restoration towards a precise and intelligent management model, effectively reduce redundant resource input and waste, improve the implementation quality and long-term operation and maintenance level of ecological restoration projects, and promote the coordinated and sustainable development of mineral resource development and regional ecological environment protection.

[0017] Other features and advantages of the invention will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the techniques described above.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating a method for constructing a coupled model for predicting the effects of mine ecological restoration, provided in an embodiment of the present invention; Figure 2 This is a schematic diagram comparing the root mean square error of an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the prediction accuracy of embodiments of the present invention; Figure 4 This is a scatter plot comparison of the prediction accuracy of an embodiment of the present invention. Figure 5 This is a schematic diagram comparing the kernel density of prediction accuracy in an embodiment of the present invention. Figure 6 This is a schematic diagram of a coupled model construction device for predicting the effect of mine ecological restoration, provided in an embodiment of the present invention. Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0022] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0023] To facilitate understanding, the method for constructing a coupled model for predicting the effect of mine ecological restoration, provided in this embodiment of the invention, will be described first, referring to... Figure 1 The method includes the following steps: Step S102: Obtain the engineering implementation parameters of the mine ecological restoration project, as well as the site condition parameters and climate environment continuous monitoring sequences corresponding to the engineering implementation parameters.

[0024] Engineering implementation parameters refer to various engineering control and implementation indicators during the implementation of mine ecological restoration projects. These include quantitative and qualitative indicators such as vegetation reconstruction methods, soil improvement ratios, landform reshaping techniques, restoration project layout density, and construction control parameters, reflecting the implementation intensity and technical approach of the restoration project. Site condition parameters refer to the inherent spatial attribute parameters of the restoration site, including non-temporally fixed parameters such as topography, soil texture, soil layer thickness, hydrogeology, slope gradient, and regional vegetation base. Continuous climate and environmental monitoring sequences are long-term climate and environmental data that can be obtained through meteorological monitoring equipment or remote sensing monitoring methods. These include continuous time-series monitoring values ​​of precipitation, temperature, humidity, sunshine duration, and frequency of extreme weather events, representing dynamically changing external influence parameters.

[0025] By acquiring engineering implementation parameters, site condition parameters, and continuous monitoring sequences of climate and environment, this invention can comprehensively capture information from three core dimensions affecting the effectiveness of mine ecological restoration: human intervention, site foundation, and external driving forces. Thus, from the perspective of the synergistic effect of multiple factors and the linkage of spatiotemporal characteristics, it can explore the long-term dynamic evolution law of mine ecological restoration, making the prediction results of the prediction model scientific and accurate.

[0026] Step S104: Obtain dynamic and static features from engineering implementation parameters, site condition parameters, and continuous climate and environmental monitoring sequences. Based on the spatial constraint information of static features and the temporal local features of dynamic features at multiple time scales, construct the climate and environmental impact coupling features of the mine ecological restoration project.

[0027] Dynamic temporal characteristics are those that exhibit regular and fluctuating changes over time, reflecting the continuous impact of climate fluctuations and engineering timing adjustments on the remediation effect. Static characteristics are those without significant temporal changes and possessing fixed spatial attributes, used to characterize the inherent spatial conditions of the remediation site and the fixed layout of the engineering project. In one implementation, parameters with temporal change attributes in the above data can be used as dynamic temporal characteristics, such as precipitation and temperature time series data in continuous climate and environmental monitoring sequences, and construction progress and phased soil improvement ratio adjustment parameters in engineering implementation parameters; spatially fixed parameters without temporal changes can be used as static characteristics, such as topographic slope and soil texture in site condition parameters, and the remediation area layout pattern and fixed construction technology in engineering implementation parameters.

[0028] Static features are characterized by spatial fixedness, inherently carrying the topographic boundaries, soil substrate, and hydrological distribution of the remediation site. Spatial constraint information refers to the inherent spatial attributes of static features that limit the range, rate of evolution, and boundaries of dynamic temporal features. Specifically, this includes the inherent attributes of static features such as spatial boundaries, topographic limitations, and soil substrate constraints, which limit the range and evolution patterns of dynamic temporal features. Temporal local features are the local variation patterns extracted at each scale (e.g., daily precipitation fluctuations, quarterly temperature changes, annual precipitation trends) after decomposing dynamic temporal features according to different time scales (e.g., daily, seasonal, annual). These features characterize the specific manifestations of dynamic temporal features across different time dimensions.

[0029] In summary, this invention relies on the inherent attributes (spatial constraint information) of static features, such as topographic boundaries, soil base constraints, and hydrological distribution range, to couple dynamic temporal features that reflect the details of climate fluctuations and engineering time-series regulation. Using spatial constraint information as the underlying limit, it constrains and deeply integrates multi-timescale local features of dynamic temporal features, deeply exploring the spatiotemporal coupling mechanism between dynamic and static features. This results in a composite coupled feature (climate and environmental influence coupled feature) that can accurately capture the evolutionary logic of restoration effects in the spatiotemporal dimension. This coupled feature clearly depicts the inherent logic of mutual constraints and influences among multiple factors such as engineering implementation, site conditions, and climate fluctuations, realistically reproducing the complex evolutionary process of mine ecological restoration effects, and ensuring that the constructed prediction model can accurately capture the long-term evolutionary patterns of restoration effects.

[0030] Step S106: Construct a training sample set for the initial neural network based on the coupling characteristics of climate and environmental impacts, and use the training sample set to train the initial neural network model to construct a prediction model for the ecological restoration effect of the mine ecological restoration project.

[0031] The initial neural network is a basic fitting and learning framework that has not been trained with samples. A suitable deep learning neural network structure can be selected based on the prediction requirements. The training sample set is a collection of samples constructed based on the aforementioned coupling characteristics of climate and environmental impacts and corresponding measured indicators of restoration effects. After constructing the training sample set, it is input into the initial neural network for model training. Through iterative learning with samples, continuous parameter optimization, and continuous correction of prediction errors, the model's ability to capture the correlation between coupling characteristics and restoration effects is gradually improved. Ultimately, a dedicated deep learning model with long-term extrapolation and dynamic prediction capabilities for mine ecological restoration effects can be obtained—that is, an ecological restoration effect prediction model.

[0032] The aforementioned coupling characteristics of climate and environmental impacts not only preserve the details of short-term changes in environmental factors, but also extract their long-term patterns such as periodicity and trends. The model can rely on the short-term details and long-term patterns contained in the coupling characteristics to accurately match the complex relationship between restoration effects and various influencing factors, accurately capture the long-term trend of restoration effects over time, the late-stage lag response, and the cumulative effect, so that the prediction results can truly reflect the actual progress and future change patterns of mine ecological restoration, and improve the accuracy and stability of restoration effect prediction.

[0033] Furthermore, based on the above embodiments, this invention also provides another method for constructing a coupled model for predicting the effects of mine ecological restoration. This method mainly describes the construction of the coupled features of the aforementioned climate and environmental influences. For the dynamic time-series climate data in step S104, the real-time impact and long-term cumulative effect of climate factors on mine ecological restoration can be correlated; static feature data contains prior knowledge of the field of mine ecological restoration and can reflect the inherent differences in different mine scenarios. This invention can arrange the precipitation and temperature monitoring data in the spatiotemporal driving factors according to the monitoring time point order to construct a time-series feature matrix. , dimension ,in, The time series length represents the total number of time points within the monitoring period. The dimension "2" corresponds to two time series variable channels, specifically precipitation and temperature. For static feature data, a static feature matrix can be constructed based on the original features (i.e., engineering parameters, site conditions, and climate monitoring sequences). .

[0034] Furthermore, the effectiveness of mine ecological restoration is influenced by both long-term and short-term climate factors, and there are complex nonlinear interactions between climate factors, engineering parameters, and site conditions. Conventional convolutional neural networks typically use fixed-scale convolutional kernels, making it difficult to simultaneously capture the evolution patterns at different time scales in climate data. Moreover, the domain knowledge inherent in static features cannot be used to dynamically adjust the attention given to features at different samples and time scales by the convolutional kernels, leading to shortcomings. This invention employs a multi-scale dynamic convolutional module. After separating and preparing climate time-series data and static features, a set of convolutional kernels at different time scales are used in parallel to extract multi-scale climate features. An attention mechanism based on static features is introduced to dynamically calculate the fusion weights of features at each scale, thereby achieving adaptive feature fusion to construct corresponding coupled features. In specific implementation, the coupled features of climate and environmental impacts of mine ecological restoration projects can be constructed through the following steps: Step S10: Use a multilayer perceptron to perform feature mapping on the spatial constraint information of static features to obtain a context vector representing the ecological response characteristics of static features.

[0035] Step S11: Based on the context vector, determine the scale attention weights corresponding to the local features of the dynamic temporal features at each time scale.

[0036] Step S12: Based on scale attention weights, perform dynamic weighted fusion processing on the temporal local features of all time scales to obtain dynamic climate fusion feature vectors.

[0037] This invention relates to a method that dynamically assigns attention weights to feature maps at different scales based on the sample-specific information contained in static features, and obtains dynamic climate fusion features through adaptive feature fusion, expressed as follows:

[0038] In the formula, Represents a dynamic climate fusion feature vector with dimension . This study characterizes the dynamic impact of climate factors on the ecological restoration of mines, integrates climate models at different time scales (such as short-term precipitation fluctuations and long-term temperature trends), and incorporates sample-specific static background information to more comprehensively reflect the complex nonlinear relationship between climate and restoration effects. This represents the global average pooling operation, which is used to perform global average pooling on the feature map of each sample in the time dimension.

[0039] Indicates the first The attention weights for each scale feature are scalars used to measure the importance of that scale feature for the current sample prediction task. Their values ​​are determined by the static features of the current sample and satisfy the following conditions: The calculation method is expressed as follows:

[0040] This represents the scale index, with values ​​ranging from... arrive ; This represents the total number of convolutional scales, i.e., the number of convolutional kernels with different time window sizes; This represents the natural exponential function. Indicates the first The query vector corresponding to each scale has a dimension of . , used to calculate similarity with the context vector of static features by dot product (i.e. the query vector in the self-attention mechanism), is a trainable parameter; express The transpose of . This represents a multilayer perceptron used for attention computation, which maps static features to context vectors to represent the static background information of the current sample. Its parameters are trainable parameters. This represents the dimension of the hidden context vector in the attention mechanism, with examples of possible values. In one implementation, the multilayer perceptron used for attention computation can employ a lightweight fully connected neural network, which incorporates a static feature matrix. (dimension) Mapped to context vector (dimension) ,in ), used to calculate attention weights, the specific structure is: input layer (accepts a static feature matrix) Dimension Hidden layer (a fully connected layer using the ReLU activation function, the output dimension can be set to...) (e.g., 128)), Output layer (linear layer, no additional activation required, output dimension) ).

[0041] It should be noted that, The term is a dot product operation, and Output dimension is For each sample, a scalar is computed; that is, for each sample, Output dimensional vector, and Vie The dot product yields a scalar, and then softmax is applied across all scales to obtain the first scalar value for a single sample. Attention weights for each scale feature.

[0042] Furthermore, by defining convolution kernels with different time window sizes, parallel one-dimensional convolution operations can be performed on the temporal feature matrix to extract feature maps at different scales (i.e., temporal local features at the corresponding time scales), as follows:

[0043] In the formula, Indicates by the first The feature maps extracted by convolutional kernels of various scales have a dimension of . Characterized on a time scale Local pattern feature mapping of time series data. This represents the number of output channels of the convolution kernel, that is, the number of channels of the feature map extracted by each convolution kernel; This indicates the length of the time series, i.e., the total number of time points within the monitoring period; Indicates the first The time window size of each convolutional kernel, i.e., the length of the convolutional kernel in the time dimension, is given by an example value. These are used to capture short-term, medium-term, and long-term climate patterns, respectively. This represents a one-dimensional convolution operation, where the convolution kernel slides along the time dimension of the input time series data to perform one-dimensional convolution on the input time series data in order to extract local temporal features; This represents the modified linear unit activation function, used to introduce nonlinearity into the feature map; Indicates the first The weight tensor of each scale convolutional kernel, with dimension . , are trainable parameters; Indicates the first The bias vectors corresponding to the convolutional kernels of each scale have dimensions of . , are trainable parameters.

[0044] Step S13: Perform feature coupling processing on the static features and dynamic climate fusion feature vector to construct the climate and environmental impact coupling features of the mine ecological restoration project.

[0045] The aforementioned dynamic climate fusion feature vector is a fusion data that is driven by static features (mine site conditions, engineering parameters, etc.) and adaptive weights. It adaptively fuses local climate time series features (covering short-term climate fluctuations and long-term climate evolution trends) at different time scales. This data can capture the dynamic correlation between climate and mining scenarios and accurately characterize the complex nonlinear interaction between climate factors and inherent differences in mining scenarios.

[0046] The effectiveness of mine ecological restoration is the result of complex coupling and nonlinear interactions between static and dynamic temporal factors, exhibiting lag and cumulative effects. Simple feature concatenation followed by direct regression is insufficient to fully characterize the interaction mechanisms among engineering measures, site conditions, and climate sequences, as well as the evolution of restoration effects over time. This invention further couples the aforementioned static features with the fused feature to construct a coupled climate and environmental impact feature. In one implementation, a bilinear interactive attention mechanism can be used to uncover the complex correlation between the two and generate a coupled enhancement feature, represented as:

[0047] In the formula, This represents the coupling enhancement feature matrix (i.e., the coupling features of climate and environment influences), with dimensions of [missing information]. It contains deep interactive information between static conditions and dynamic climate factors through attention mechanisms. For example, the negative synergistic effect of "heavy precipitation" climate events under "high slope" site conditions will be enhanced and characterized. This represents the Sigmoid activation function, used to map attention scores to... The interval serves as a soft selection and gating mechanism; This represents the weight matrix used to generate the query vector, with dimensions [missing information]. , are trainable parameters; This represents the weight matrix used to generate the key vectors, with dimensions of . , are trainable parameters; This represents the weight matrix used to generate the value vector, with dimensions of . , are trainable parameters; This represents the Hadamard product, which is an element-wise multiplication operation. The dimension of the coupled feature space is represented by a hyperparameter, with an example value of 256. This indicates an operation that concatenates along the feature dimension; The bias vector representing the attention score calculation has dimensions of . When added, broadcast to A matrix is ​​a set of trainable parameters.

[0048] Furthermore, after constructing a training sample set using the aforementioned coupling features, the model is built. In this embodiment of the invention, a neural network prediction model based on coupled attention and time-varying memory is adopted. The high-order nonlinear interaction between static features and dynamic climate features can be modeled through a feature cross-enhancement module (i.e., step S13 above). A time-varying decay memory gate can be further introduced to simulate the accumulation and decay dynamics of the repair effect. Combined with derived engineering and climate matching features, multi-step dynamic prediction values ​​are generated through a constrained output layer. The error between this prediction value and the actual monitored value can be used to determine the model training accuracy, thereby guiding iterative optimization of the model.

[0049] In practical implementation, for step S106 above, an ecological restoration effect prediction model can be constructed through the following steps: Step S20: Based on the restoration effect index corresponding to the climate and environmental impact coupling characteristics, construct a training sample set corresponding to the climate and environmental impact coupling characteristics.

[0050] The remediation effectiveness indicators can be determined based on multi-source heterogeneous data from historical remediation projects (such as engineering parameters, site conditions, and climate monitoring sequences). Corresponding to the above embodiment, engineering parameters may include the type of remediation project (such as tree planting, grass planting, soil improvement, and slope protection), project duration (months), material usage (tons / hectare), and planting density (plants / hectare). Site conditions may include soil pH (dimensionless), soil organic matter content (%), slope (degrees), altitude (meters), soil thickness (centimeters), site type (such as mountains, hills, plains, and tidal flats), soil type (such as sandy soil, clay, loam, and gravelly soil), and vegetation species (such as trees, shrubs, herbs, and mosses). Climate monitoring sequences may include precipitation sequences (millimeters) and temperature sequences (degrees Celsius). These data are continuously collected at monitoring points (such as monthly or daily) to form a spatiotemporal sequence. Correspondingly, the remediation effectiveness indicators, as prediction targets, include vegetation cover (normalized percentage) and soil organic matter content (normalized percentage), which can be determined through multi-period dynamic monitoring. The above data can be obtained through monitoring reports, sensor networks, remote sensing imagery, and field surveys of mine restoration projects, ensuring spatiotemporal consistency and accuracy. Data annotation can be based on actual measurements, quantitatively labeling the restoration effect indicators for each sample at each predicted time step. The annotation category is continuous numerical, namely the normalized values ​​(range) of vegetation cover and soil organic matter content. Furthermore, during dataset construction, the collected raw data can be divided into samples, with each sample corresponding to a historical restoration project, including its feature data (engineering, site, climate) and effect indicator labels for multiple time steps. Subsequently, it is divided into training set, validation set, and test set for model training, hyperparameter tuning, and performance evaluation.

[0051] Step S21: Using a time-varying decay memory gate, determine the memory state vector of the repair effect corresponding to the coupling characteristics of climate and environment influence.

[0052] The effectiveness of mine ecological restoration depends not only on current conditions but also on historical conditions, and the effect naturally decays or strengthens over time. This invention uses a time-varying decay memory gate to simulate this dynamic process, obtaining a memory state vector to adjust the predicted time-series trajectory, expressed as:

[0053] In the formula, Indicates the first The memory state vector for each prediction time step has a dimension of [missing information]. This state vector encodes the "historical memory" of the repair effect up to the current time step, and the memory state vector of the initial time step. Set it to the zero vector. Indicates the first The memory state vector of each prediction time step; This represents the index of the prediction time step, with a value range of 1. For example, it can be used as a unit of month to predict the future number of months. The repair effect after one month; The dimension of the memory state vector is a hyperparameter, with an example value of 128. This indicates the total prediction step size that the model needs to output, for example, predicting the next 36 months. This represents the weight matrix in the memory update transformation, with dimension 1. , are trainable parameters; This represents the bias vector in the memory update transformation, with dimension . , are trainable parameters; This represents the hyperbolic tangent activation function, which constrains the output of the memory update transform to... Interval.

[0054] This represents the time-varying decay rate function, with a range of . This value controls the proportion of historical memory states that are retained, and it varies with time intervals. Increase and decrease, the calculation method is expressed as:

[0055] In the formula, The time interval is the time interval between the current prediction time step and the reference starting point, in months, and is a scalar constant that is the same for all samples in this prediction step. The set of parameters representing the decay rate function, i.e. , This represents the linear coefficient in the decay rate function. The coefficients of the quadratic term in the decay rate function are all trainable parameters; This represents the first minimal positive integer, used to prevent numerical underflow during exponentiation calculations. An example value is shown below. ; This represents the function that takes the maximum value.

[0056] Step S22: Train the initial neural network model based on the repair effect memory state vector, training sample set, and preset engineering and climate matching features.

[0057] pass This represents the characteristics of the matching degree between engineering and climate, with dimensions of 1. , is a scalar characteristic, and the closer its value is to 1, the more adapted the engineering measures are to the expected local climate conditions. The calculation method is expressed as follows:

[0058] In the formula, The engineering feature subset represents a subset of engineering features selected from the original input, with a dimension of [dimensional value missing]. ; This represents the first deterministic mapping function used to map a subset of engineering features to a scalar representing the "strength of an engineering material or method". This represents the second deterministic mapping function, used to map climate statistics to a scalar representing the "intensity of climate stress". This represents the maximum possible difference value used for normalized matching degree calculation, and is a preset constant.

[0059] In one implementation, The preset method is achieved through historical data statistics, specifically by calculating the historical samples. The maximum absolute value of the difference is multiplied by a preset safety factor (e.g., 1.2) as the result. In another implementation, a preset can be estimated based on domain knowledge. Specifically, based on the possible range of engineering intensity and climate stress, a theoretically maximum possible difference value can be estimated by human experience. The aim is to ensure that, in the vast majority of cases, The calculation result falls within Within the interval. In one implementation, This includes features directly related to the restoration project, such as: restoration project type (discrete), material usage (continuous, tons / hectare), project duration (continuous, months), and planting density (continuous, plants / hectare). Furthermore, it can be expanded based on domain knowledge. It should be noted that during the calculation of the project-climate matching characteristics, The item maps engineering characteristics to a "engineering strength" scalar. This term maps climate statistics to a scalar of "climate stress intensity." The smaller the difference between the two, the more adaptable the engineering measures are to the local climate. The closer it is to 1, the greater the difference. The closer to 0, and Used for normalization, ensuring The range of values ​​is .

[0060] For the first deterministic mapping function mentioned above, a preset linear weighted sum calculation method can be used for mapping. For example, let... Including material usage Project duration Planting density Continuous features (converted to raw dimensional values), and embedded scalar representations of repair project types. (For example, mapping categories to preset intensity coefficients: tree planting = 0.8, grass planting = 0.6, soil improvement = 0.9, slope protection = 0.7), then the calculation method of the first deterministic mapping function is expressed as follows:

[0061] The values ​​0.1, 0.3, 0.4, and 0.2 are weights preset based on domain knowledge.

[0062] For the second deterministic mapping function mentioned above, a preset linear weighted sum calculation method can be used for mapping. For example, let... Includes the expected average precipitation during the forecast period (mm) and expected average temperature (in degrees Celsius), the second deterministic mapping function is calculated as follows:

[0063] The values ​​of 0.4 and 0.6 are preset weights based on domain knowledge. Indicates the forecast period The expected climate conditions statistics within the forecast period include key climate statistics such as: 1) annual average precipitation (mm), 2) annual average temperature (°C), 3) precipitation variability (standard deviation, mm), and 4) temperature variability (standard deviation,°C). The specific characteristics to be selected need to be pre-selected based on the climate characteristics of the mine location and the impact mechanism on the ecological restoration process.

[0064] Furthermore, the coupling features of climate and environmental impacts, the memory state vector of the last prediction time step, and the matching degree features between engineering and climate are concatenated and mapped to a prediction sequence through a fully connected output layer (each prediction target is independent). A truncation function is used to ensure that the predicted values ​​fall within the specified range. Within the interval, the normalized multi-step dynamic prediction value is obtained, expressed as:

[0065] In the formula, Indicates that all samples are in the first... In terms of individual repair efficacy indicators, the future A matrix consisting of the predicted values ​​at each time step, with dimensions of [missing information]. ; The index represents the target indicator of the repair effect, and its value range is... ; This represents the total number of predicted targets, for example, the value... This indicates that the model outputs two indicators: vegetation cover and soil organic matter content. This function represents a truncation function that forces the output value to be limited to a specific value. Within the interval, corresponding to the theoretical range of the normalized ecological restoration effect index, the physical rationality of the prediction results is ensured; Indicates the first The weight matrix of the prediction target output layer has dimensions of . , are trainable parameters used to linearly map a comprehensive feature vector, which integrates static features, dynamic climate features, historical memory, and engineering climate matching, to the future. The predicted value at each time step; Indicates the first The bias vectors of the prediction target output layer, with dimension , are . , are trainable parameters. This indicates the last prediction time step (i.e., the 1st prediction time step). The memory state vector of step (step), with dimension ( ). Its encoding is the cumulative state of the repair effect as of the end of the prediction period, which is affected by all historical and current conditions, and is used to provide dynamic evolution information for the final prediction.

[0066] Step S23: Based on the prediction uncertainty of model training, adjust the loss weights of the initial neural network to construct a prediction model for the ecological restoration effect of the mine ecological restoration project.

[0067] The model iterative training and parameter update in this embodiment of the invention are an optimization process based on the constructed dataset and the defined multi-objective adaptive weighted loss function. At the start of training, all trainable parameters are initialized, and an optimization algorithm (such as the Adam optimizer) is used to minimize the total loss function. In each training cycle, forward propagation calculates the normalized feature matrix. Dynamic climate integration characteristics Enhanced coupling features Memory state vector and predicted value Then, the total loss function is calculated, which incorporates multi-objective adaptive weights and temporal decay weights to balance the importance of different prediction objectives (such as vegetation cover and soil organic matter content) and different time steps; the gradient is calculated through backpropagation, and all trainable parameters are updated through the optimizer.

[0068] After each training cycle, update the prediction uncertainty. This is used to dynamically adjust the weights of the loss function in the next cycle, thereby optimizing the balance of multi-task learning. The criterion for stopping iteration is based on validation set performance monitoring: when the validation loss no longer decreases or begins to increase within a preset number of consecutive cycles (e.g., 10 cycles), an early stopping mechanism is triggered to terminate training and prevent overfitting; at the same time, a maximum number of training cycles is set as a backup stopping condition. After training is completed, the model parameters with the minimum validation loss are saved for subsequent prediction tasks.

[0069] In mine ecological restoration, predicting the effectiveness typically requires simultaneous dynamic prediction of multiple ecological indicators. These indicators have different dimensions, and their importance may change with the restoration stage at different prediction time steps. Furthermore, the learning difficulty and uncertainty of each prediction objective vary. Using an equally weighted loss function can lead to a model biased towards learning simpler objectives while struggling to optimize those with high uncertainty, thus affecting the overall accuracy of multi-objective prediction. This invention employs a loss function based on multi-objective adaptive weighted prediction. It not only assigns adaptive weights to each prediction objective to balance the learning difficulty of different objectives but also introduces time-decaying weights to emphasize the importance of recent prediction accuracy. By dynamically adjusting the contribution of each objective's loss term, it guides the model to achieve a better equilibrium state in multi-task learning, expressed as:

[0070] In the formula, This represents the total loss function. By minimizing the loss function, the model parameters are optimized to make the multi-objective predictions approximate the true values ​​and to balance the importance of different objectives and time steps. This represents the Charbonnier loss function, a robust and smooth variant of the L1 loss that is insensitive to outliers in the training data. Indicates the first The temporal decay weight for each prediction time step, its value as... The exponential decay of the sum of its parts (increases and decreases) is used to give higher priority to recent forecasts in loss calculations. The calculation method is expressed as follows: ; Indicates the first The sample at the th The first future prediction time step The true value of each performance indicator, which has been pre-normalized; Let represent the predicted value of the r-th performance index for the i-th sample at the t-th prediction time step. The Line number Column elements.

[0071] This represents the second smallest positive constant, used to prevent the denominator of the formula from being zero. Examples of its values ​​are shown below. ; Indicates the first At the end of the training cycle, the estimated first... The prediction uncertainty of a target, the target with high uncertainty ( Large), loss weight Small, to avoid the model focusing too much on difficult-to-predict targets, thus balancing multi-task learning, the calculation method is expressed as:

[0072] In the formula, This represents the exponential smoothing factor, used to balance historical values ​​with current observations when iteratively updating uncertainty estimates; its value lies between... and Between, the value examples are as follows . This indicates the operation of calculating the standard deviation of all elements in a given set. Indicates the first At the end of the training cycle, the estimated first... The prediction uncertainty of the first prediction target, the second Initial value of the prediction uncertainty of each target Set as . This represents the training cycle index. Indicates the first One training cycle, all All samples In the prediction step, the first The set of Charbonnier loss values ​​for each target. Indicates the sample index.

[0073] In practical implementation, These are the true values ​​provided in the training data. Before training the model, it is necessary to collect monitoring data from multiple phases of historical mine ecological restoration projects. For example, for each historical restoration project (such as the first phase...), the data should be collected. (one sample), which needs to be repaired. Months (corresponding to the predicted time step) vegetation coverage () ), soil organic matter content ( The actual measured values ​​of the indicators constitute Furthermore, in order to ensure stable training and efficient loss function calculation, It needs to be normalized to the same Interval.

[0074] Furthermore, mine ecological restoration data exhibits typical multi-source heterogeneity, encompassing continuous numerical data, discrete encoded data, and spatiotemporal sequences. Due to the significant differences in the dimensions and scales of each feature, and the existence of physical correlations and domain knowledge among the features, this invention designs normalization strategies for the three types of heterogeneous features—continuous, discrete, and spatiotemporal sequences—to construct a normalized feature matrix of the original data. This matrix can also be used as a static feature matrix, corresponding to normalized continuous features, discrete feature embedding vectors, and spatiotemporal sequence statistical features, with dimensions of [missing information]. .

[0075] Conventional normalization methods struggle to effectively handle both discrete and continuous features simultaneously, and fail to incorporate physical relationships between features and expert prior knowledge, resulting in a mismatch between the contribution of normalized features to the model output and their physical importance. This invention addresses this by introducing physical constraint weights and adaptive learnable parameters, using an adaptive normalization method based on physical constraints to uniformly map features of different dimensions to a dimensionless interval.

[0076] In practical implementation, this invention, based on the data characteristics of engineering implementation parameters, site condition parameters, and continuous climate and environmental monitoring sequences, divides these data into multiple sets of data to be processed. Furthermore, each set of data is normalized to construct a standardized feature matrix corresponding to the mine ecological restoration project.

[0077] In one implementation, the data characteristics include: continuous numerical characteristics, classification category characteristics, and monitoring time-series characteristics; multiple sets of data to be processed include continuous parameters corresponding to continuous numerical characteristics, discrete classification parameters corresponding to classification category characteristics, and climate monitoring data corresponding to monitoring time-series characteristics. Correspondingly, each set of data can be normalized using the following steps: 1) For continuous parameters, this embodiment of the invention performs adaptive physical constraint normalization processing on the corresponding continuous parameters based on the data distribution of the continuous parameters and the degree of physical influence of the continuous parameters on the ecological restoration effect, and generates the first preprocessed data.

[0078] For continuous features, this embodiment of the invention adopts a normalization framework based on the physical range of the features, and introduces an adaptive scaling factor and physical constraint weights for adjustment, as follows:

[0079] In the formula, Indicates the first The normalized values ​​of continuous features are constrained to a range of 1 / 2. This is used to eliminate dimensional differences and preserve the physical comparability of the feature value between different samples. Indicates the first The original values ​​of each continuous feature are the original continuous data input to the model. This represents the function that takes the minimum value. This represents the function that takes the maximum value. This represents the hyperbolic tangent function, used to strictly constrain the normalized values ​​of continuous features to a certain value. And it does not affect the relative comparability of the feature distributions. Indicates the first An adaptive scaling factor for a continuous feature is a trainable scalar parameter used to dynamically adjust the normalized output scale of the feature based on the data distribution and task objectives. Indicates the first The physical constraint weights of continuous features are scalar constants preset based on the known physical impact of the feature on the ecological restoration effect. These weights are used to incorporate domain knowledge into feature preprocessing to adjust the initial importance of different features. For example, features with a significant impact on the restoration effect (such as project duration) might be preset to a certain value. The default setting affects general characteristics (such as planting density). .

[0080] In one embodiment, continuous characteristics include: project duration in months (representing the duration of the remediation project); material usage in tons per hectare (representing the mass of remediation material used per hectare); soil pH value, dimensionless (representing soil acidity / alkalinity); soil organic matter content in % (representing the percentage of organic matter by mass in the soil); slope in degrees (representing the angle of terrain slope); altitude in meters (representing altitude); soil thickness in centimeters (representing soil layer thickness); and planting density in plants per hectare (representing the number of plants planted per hectare). The original values ​​of continuous features It can refer to any of the above characteristics.

[0081] 2) For discrete classification data, norm normalization can be performed on the corresponding discrete classification parameters based on the one-hot encoding vector corresponding to the discrete classification parameters to generate the second preprocessed data.

[0082] One-hot encoded vector representations transform discrete categorical variables, such as non-continuous, non-numerical features like mine site conditions and engineering implementation methods, into vector forms containing only 0s and 1s. Furthermore, this can be mapped to a continuous space through a learnable embedding layer and normalized using norm, resulting in:

[0083] In the formula, Indicates the first The normalized embedding vector of each discrete feature (i.e., the second preprocessed data) has its magnitude roughly constrained to... Near the interval, it is used to transform discrete category information into a comparable continuous vector representation. Indicates the first The continuous embedding vector obtained after mapping discrete features is used to transform discrete category information into a dense, computable continuous representation. The calculation method is expressed as follows: . Indicates the first The one-hot encoded vector of each discrete feature is a binary vector representation, and its length is... Equal to the total number of categories for that feature, only at the corresponding category index. The rest are ; Indicates the first The embedding matrix corresponding to each discrete feature is a trainable parameter with dimension . This is used to map discrete categories to a continuous one. In a 3D vector space; Indicates the first The total number of categories for each discrete feature; Indicates the first The preset dimension of each discrete feature embedding vector, with examples of possible values. ; This represents the L2 norm.

[0084] In one embodiment, discrete features include: remediation project type, with categories including tree planting, grass planting, soil improvement, and slope protection, totaling 4 categories; site type, with categories including mountains, hills, plains, and tidal flats, totaling 4 categories; vegetation type, with categories including trees, shrubs, herbs, and mosses, totaling 4 categories; and soil type, with categories including sandy soil, clay, loam, and gravelly soil, totaling 4 categories. For the first... A discrete feature, It is a length of The one-hot encoded vector, for example, if the "repair project type" is "tree planting", then It should be noted that, since the L2 norm normalized components are usually no more than 1, the calculation method for the normalized embedding vector aims to make the embedding vector components roughly distributed within a certain range. In practice, this can be adjusted through training.

[0085] 3) For climate monitoring data corresponding to the monitoring time series characteristics, adaptive physical constraint normalization can be performed on the climate monitoring data based on the continuous characteristic statistical parameters of the climate monitoring data to generate third preprocessed data.

[0086] Climate monitoring data can also be characterized as spatiotemporal series data. In this embodiment of the invention, long series are not processed directly. Instead, they are first compressed into representative statistical features, and then these statistics are normalized as continuous features, as follows:

[0087] In the formula, Indicates the first The mean statistic of a spatiotemporal series represents the overall level of the series during the monitoring period; Indicates the first The standard deviation statistic of a spatiotemporal series represents the degree of fluctuation of the series during the monitoring period; Indicates the first The maximum value statistic of a spatiotemporal sequence, which characterizes the peak level of the sequence during the monitoring period; Indicates the first The minimum value statistic of a spatiotemporal sequence represents the valley level of the sequence during the monitoring period. An adaptive physical constraint normalization function representing continuous characteristics. Indicates the first Mean statistic of a spatiotemporal series The value after processing by the adaptive physical constraint normalization method for continuous features; Indicates the first Standard deviation statistic of a spatiotemporal series The value after processing by the adaptive physical constraint normalization method for continuous features; Indicates the first Maximum value statistics of a spatiotemporal sequence The value after processing by the adaptive physical constraint normalization method for continuous features; Indicates the first Minimum statistic of a spatiotemporal sequence The value after processing by the adaptive physical constraint normalization method for continuous features.

[0088] In one embodiment, the spatiotemporal sequence characteristics refer to climate data collected in time sequence during the monitoring period, specifically including: precipitation sequence, in millimeters, representing monthly or daily precipitation monitoring values; and temperature sequence. The unit is degrees Celsius, representing the average monthly or daily temperature monitoring value. The mean is extracted for each series. Standard deviation Maximum value Minimum value Four statistics are used as continuous features for normalization. For example, for a precipitation series, This indicates the average precipitation during the monitoring period.

[0089] 4) Further, the first preprocessed data, the second preprocessed data, and the third preprocessed data are spliced ​​together to construct a standardized feature matrix corresponding to the mine ecological restoration project.

[0090] In summary, all normalized continuous features, discrete feature embedding vectors, and spatiotemporal sequence statistical features can be concatenated to form a normalized feature matrix. , dimension ;in, This indicates the total number of data samples. This represents the total feature dimension after concatenating all features.

[0091] In summary, the embodiments of the present invention are innovative compared to the prior art in the following aspects: 1. An adaptive normalization method based on physical constraints is adopted. By introducing a learnable scaling factor and preset physical weights, it can uniformly process multi-source heterogeneous data such as continuous, discrete and spatiotemporal sequences. This not only eliminates the difference in units, but also embeds domain knowledge into feature preprocessing, thereby improving the physical comparability of features and the quality of model input.

[0092] 2. A multi-scale dynamic convolutional fusion module is adopted, which uses convolutional kernels with different time windows to extract climate sequence features in parallel, and uses attention weights generated by static features for dynamic fusion. This realizes adaptive multi-scale climate pattern mining guided by static background information, which enhances the pertinence and interpretability of feature extraction.

[0093] 3. The prediction mechanism of coupled attention and time-varying memory is adopted. The high-order nonlinear coupling of static and dynamic features is modeled by bilinear interactive attention, and the accumulation and lag effects of the repair effect are simulated by time-varying decay memory gate. Combined with the explicitly constructed engineering-climate matching degree feature, the model can deeply characterize the complex evolution dynamics of the repair effect.

[0094] 4. A multi-objective adaptive weighted loss function is adopted. By dynamically estimating the uncertainty of each prediction objective, the loss weight is adaptively allocated. A time-series decay weight is introduced to emphasize recent predictions. This effectively balances the differences in difficulty and importance of different indicators and different time steps in multi-objective learning, and guides the model to achieve more balanced optimization.

[0095] Furthermore, the trained coupling model can be used to conduct a multi-step prospective evaluation of new restoration projects. Correspondingly, characteristic data of the new project can be collected first, including engineering parameters (such as restoration project type, project duration, material usage, and planting density), site conditions (such as soil pH, soil organic matter content, slope, altitude, soil thickness, site type, soil type, and vegetation type), and climate data (such as historical precipitation sequences, temperature sequences, or expected climate statistics for the forecast period). Furthermore, adaptive normalization and physical constraint fusion can be performed on the features to form a corresponding normalized feature matrix. For example, continuous features (such as project duration, material usage, soil pH, soil organic matter content, slope, altitude, soil thickness, and planting density) are converted into normalized values; discrete features (such as restoration project type, site type, vegetation type, and soil type) are converted into embedding vectors; and spatiotemporal sequence features (such as precipitation and temperature sequences) are extracted and normalized by extracting statistics (mean, standard deviation, maximum, and minimum values). Furthermore, the corresponding data can be input into a pre-trained coupled model, and coupled enhanced features can be generated through coupled attention and time-varying memory modules. and memory state vector And combined with the characteristics of engineering and climate matching Ultimately, the future is generated by constraining the output layer. Predicted repair efficacy at each time step (e.g., 36 months) (Each target) (Corresponding to a matrix). The predicted value is the normalized value (range). The output can be denormalized to its original dimensions (such as percentage of vegetation cover or percentage of soil organic matter content) as needed. This output provides a dynamic trajectory of the restoration effect over time, enabling decision-makers to assess the long-term effectiveness of restoration programs, optimize engineering adjustment strategies, and achieve intelligent management of the ecological restoration process.

[0096] In one embodiment, this invention also evaluates the comprehensive performance of different prediction models on the task of dynamically predicting vegetation cover in mine ecological restoration, comparing the capabilities of the proposed coupled model with conventional techniques in processing multi-source heterogeneous data and predicting long-term ecological effects. The models compared include linear regression, random forest, support vector machine, long short-term memory network model, convolutional neural network model, and Transformer model. These models represent the mainstream technical paths from traditional statistical methods to modern deep learning. Linear regression and random forest, as classic machine learning methods, are widely used for ecological environment prediction; support vector machines have advantages in processing high-dimensional data; long short-term memory network model, convolutional neural network model, and Transformer model are deep learning architectures that have shown outstanding performance in time series prediction and feature extraction in recent years. All compared models are trained and tested using the same dataset as the model of this invention to ensure the fairness of the comparison.

[0097] Reference Figure 2 , Figure 3The diagrams show the root mean square error (RMSE) and prediction accuracy of different methods. The experimental results are presented in two subplots: a box plot shows the distribution of prediction errors for each model, with the ordinate representing the dimensionless normalized error value (RMSE), indicating higher accuracy; and a line plot visually compares the degree of fit between the predicted vegetation cover trajectories of different models and the actual values ​​over the next 36 months, with the ordinate representing vegetation cover using normalized values. The box plot clearly shows that the coupled model of this invention has the lowest error box position, the shortest length, and fewer outliers, indicating the highest prediction accuracy and most stable results. Furthermore, the line plot shows that the prediction curve of this invention almost overlaps with the black solid line representing the actual value, especially in the later stages of prediction, where other models' prediction curves show significant deviations or fluctuations, while this invention's model maintains a high degree of consistency. The experimental results demonstrate that this invention, through adaptive normalization and physical constraint fusion, multi-scale temporal feature mining, and coupled attention mechanisms, can more accurately capture the complex evolution of restoration effects over time, significantly improving the reliability of long-term predictions.

[0098] Furthermore, this embodiment of the invention also evaluates the differences in prediction accuracy among various models under different types of restoration projects, and examines the adaptability and predictive ability of the models to different human intervention measures. The comparison models include linear regression, random forest, long short-term memory network models, and the coupled model of this invention. The experimental configuration covers four typical mine restoration projects: tree planting, grass planting, soil improvement, and slope protection. The prediction accuracy of each model under different project types is statistically analyzed. This accuracy is a dimensionless ratio, and a higher value is better. (Refer to...) Figure 4 , Figure 5 The experimental results are presented through scatter plots and kernel density estimation plots: the ordinate of the scatter plot represents the prediction accuracy, where each point represents the prediction result of a sample, and the horizontal line segment represents the average value of the data set; the abscissa of the kernel density plot represents the accuracy, and the ordinate represents the probability density. The scatter plot shows that for each type of repair project, the orange diamond-shaped scatter plots representing the model of this invention cluster in a higher ordinate region, and its average level is also significantly higher than other models. The kernel density plot further confirms the advantages of this technology from an overall distribution perspective. The density curve of the model of this invention has the highest peak and is located furthest to the right, meaning that its prediction accuracy not only has a high average level, but also that most samples are concentrated in this high-precision range, demonstrating excellent stability.

[0099] Experimental results show that the introduction of engineering and climate matching features and the feature cross-enhancement module in this invention enable the model to accurately understand the intrinsic mechanisms of different engineering measures, thereby making accurate dynamic predictions under various remediation schemes.

[0100] Based on the above embodiments, this invention also provides a coupled model construction device for predicting the effects of mine ecological restoration, referring to... Figure 6 The device includes: a data acquisition module 10, used to acquire engineering implementation parameters of the mine ecological restoration project, as well as site condition parameters and continuous climate and environmental monitoring sequences corresponding to the engineering implementation parameters; a data processing module 20, used to acquire dynamic temporal features and static features from the engineering implementation parameters, site condition parameters, and continuous climate and environmental monitoring sequences, and to construct the climate and environmental impact coupling features of the mine ecological restoration project based on the spatial constraint information of the static features and the temporal local features of the dynamic temporal features at multiple time scales; and an execution module 30, used to construct a training sample set for an initial neural network based on the climate and environmental impact coupling features, and to train the initial neural network using the training sample set to construct a prediction model for the ecological restoration effect of the mine ecological restoration project.

[0101] The coupled model construction device for predicting the effect of mine ecological restoration provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0102] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described... Figure 1 The steps of the method are shown. Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the above-described steps. Figure 1 The steps of the method are shown. This invention also provides a schematic diagram of the structure of an electronic device, as shown. Figure 7 The diagram shows the structure of the electronic device, which includes a processor 101 and a memory 100. The memory 100 stores computer-executable instructions that can be executed by the processor 101. The processor 101 executes the computer-executable instructions to implement the above-mentioned... Figure 1 The method shown.

[0103] exist Figure 7In the illustrated embodiment, the electronic device further includes a bus 102 and a communication interface 103, wherein the processor 101, the communication interface 103, and the memory 100 are connected via the bus 102. The memory 100 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk drive. Communication between this system network element and at least one other network element is achieved through at least one communication interface 103 (which can be wired or wireless), using the Internet, wide area network, local area network, metropolitan area network, etc. Bus 102 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, or an AMBA (Advanced Microcontroller Bus Architecture) bus. AMBA defines three types of buses: APB (Advanced Peripheral Bus), AHB (Advanced High-performance Bus), and AXI (Advanced eXtensible Interface). Bus 102 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 7The diagram uses only a single double-headed arrow, but this does not imply a single bus or a single type of bus. Processor 101 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 101 or by instructions in software form. Processor 101 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor 101 reads information from the memory and, in conjunction with its hardware, completes the aforementioned tasks. Figure 1 The method shown.

[0104] The computer program product provided by this invention, which is a method and apparatus for constructing a coupled model for predicting the effect of mine ecological restoration, includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the preceding method embodiments, which will not be repeated here. If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0105] In the description of this invention, it should be noted that the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Finally, it should be noted that the above embodiments are merely specific implementations of this invention, used to illustrate the technical solutions of this invention, and not to limit it. The scope of protection of this invention is not limited thereto. Although this invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in this invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention, and should all be covered within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A method for constructing a coupled model for predicting the effects of mine ecological restoration, characterized in that, The method includes: Obtain the engineering implementation parameters of the mine ecological restoration project, as well as the site condition parameters and climate environment continuous monitoring sequences corresponding to the engineering implementation parameters; Dynamic and static features are obtained from the engineering implementation parameters, site condition parameters, and continuous climate and environment monitoring sequences. Based on the spatial constraint information of the static features and the temporal local features of the dynamic features at multiple time scales, the climate and environment impact coupling features of the mine ecological restoration project are constructed. A training sample set for an initial neural network is constructed based on the coupling characteristics of the climate and environmental impacts. The initial neural network is then trained using the training sample set to construct a predictive model for the ecological restoration effect of the mine ecological restoration project.

2. The method according to claim 1, characterized in that, The steps for constructing the climate and environmental impact coupling characteristics of the mine ecological restoration project based on the spatial constraint information of the static features and the temporal local features of the dynamic temporal features at multiple time scales include: A multilayer perceptron is used to perform feature mapping on the spatial constraint information of the static features to obtain a context vector characterizing the ecological response characteristics of the static features; Based on the context vector, the scale attention weights corresponding to the temporal local features of the dynamic temporal features at each time scale are determined; the scale attention weights are used to characterize the degree of influence of the temporal local features at the corresponding time scale on the ecological restoration effect; Based on the scale attention weights, the temporal local features of all time scales are dynamically weighted and fused to obtain a dynamic climate fusion feature vector. The static features and the dynamic climate fusion feature vector are coupled to construct the climate and environmental impact coupling features of the mine ecological restoration project.

3. The method according to claim 2, characterized in that, Based on the context vector, the step of determining the scale attention weights corresponding to the local temporal features of the dynamic temporal features at each time scale includes: Based on the dot product similarity between the temporal local features at each time scale and the context vector, the scale attention weight of the temporal local features at the corresponding time scale is determined.

4. The method according to claim 2, characterized in that, The steps of coupling the static features with the dynamic climate fusion feature vector to construct the climate and environmental impact coupling features of the mine ecological restoration project include: Based on the bilinear interactive attention mechanism, the static features and the dynamic climate fusion feature vector are subjected to bilinear interactive coupling processing to construct the climate and environmental impact coupling features of the mine ecological restoration project.

5. The method according to claim 1, characterized in that, The steps of constructing a training sample set for an initial neural network based on the aforementioned climate and environmental impact coupling characteristics, and using the training sample set to train the initial neural network to construct a prediction model for the ecological restoration effect corresponding to the mine ecological restoration project, include: Based on the restoration effect index corresponding to the climate and environmental impact coupling characteristics, a training sample set corresponding to the climate and environmental impact coupling characteristics is constructed. A time-varying decay memory gate is used to determine the repair effect memory state vector corresponding to the coupled features of the climate and environment influences; The initial neural network is trained based on the repair effect memory state vector, the training sample set, and the preset engineering and climate matching degree features. Based on the prediction uncertainty of the model training, the loss weights of the initial neural network are adjusted to construct a prediction model for the ecological restoration effect of the mine ecological restoration project.

6. The method according to claim 5, characterized in that, The initial neural network is trained based on a preset total loss function, which is constructed based on the Charbonnier loss function and temporal decay weights. The total loss function includes multiple repair effect prediction targets, and the loss weight of each repair effect prediction target is calculated based on the corresponding prediction uncertainty.

7. The method according to claim 1, characterized in that, The method further includes: Based on the data characteristics of the project implementation parameters, the site condition parameters, and the continuous climate and environment monitoring sequence, the project implementation parameters, site condition parameters, and continuous climate and environment monitoring sequence are divided into multiple groups of data to be processed. Each set of data to be processed is subjected to data normalization to construct a standardized feature matrix corresponding to the mine ecological restoration project.

8. The method according to claim 7, characterized in that, The data characteristics include: continuous numerical characteristics, classification characteristics, and monitoring time-series characteristics; the multiple sets of data to be processed include continuous parameters corresponding to the continuous numerical characteristics, discrete classification parameters corresponding to the classification characteristics, and climate monitoring data corresponding to the monitoring time-series characteristics. The steps of performing data normalization processing on each group of data to be processed and constructing a standardized feature matrix corresponding to the mine ecological restoration project include: Based on the data distribution of the continuous parameters and the degree of physical influence of the continuous parameters on the ecological restoration effect, the corresponding continuous parameters are subjected to adaptive physical constraint normalization to generate the first preprocessed data. Based on the one-hot encoding vector corresponding to the discrete classification parameters, norm normalization is performed on the corresponding discrete classification parameters to generate the second preprocessed data. Based on the continuous characteristic statistical parameters of the climate monitoring data, adaptive physical constraint normalization is performed on the climate monitoring data to generate third preprocessed data. The first preprocessed data, the second preprocessed data, and the third preprocessed data are concatenated to construct a standardized feature matrix corresponding to the mine ecological restoration project.

9. A method for dynamically predicting the effects of mine ecological restoration, characterized in that, The method includes: The method involves acquiring data on a mine ecological restoration project to be predicted, inputting the data into a pre-constructed ecological restoration effect prediction model, and outputting a dynamic prediction result of the restoration effect of the mine ecological restoration project data. The ecological restoration effect prediction model is constructed based on the coupled model construction method for predicting the effect of mine ecological restoration as described in any one of claims 1-8. Based on the dynamic prediction results of the restoration effect, an ecological restoration project plan for the target mine is determined.

10. A coupled model construction device for predicting the effects of mine ecological restoration, characterized in that, The device includes: The data acquisition module is used to acquire the engineering implementation parameters of the mine ecological restoration project, as well as the site condition parameters and climate environment continuous monitoring sequences corresponding to the engineering implementation parameters; The data processing module is used to obtain dynamic temporal features and static features from the engineering implementation parameters, the site condition parameters and the continuous climate and environment monitoring sequence, and to construct the climate and environment impact coupling features of the mine ecological restoration project based on the spatial constraint information of the static features and the temporal local features of the dynamic temporal features at multiple time scales. The execution module is used to construct a training sample set for an initial neural network based on the coupled characteristics of the climate and environmental impacts, and to train the initial neural network using the training sample set to construct a prediction model for the ecological restoration effect of the mine ecological restoration project.

Images