Preparation and Construction Method of High-Strength Impermeable Coal-Based Solid Waste Polymer Grouting Material for Underground Goaf Filling

Abstract

This invention discloses a preparation and construction method for a high-strength, impermeable coal-based solid waste geopolymer grouting material for filling underground goaf areas, belonging to the technical field of underground engineering filling materials. The method includes three steps: material composition design, core preparation process, and AI dynamic proportion optimization. Using coal gangue and other solid waste as raw materials, component A is prepared through a cascade mechanical activation process, combined with a composite activator component B. A lightweight neural network model is constructed based on the DEGI-BPNN algorithm, and the ratio of components A and B is dynamically optimized in conjunction with underground hydrogeological conditions. This invention solves the problems of high carbon emissions, low solid waste activity, difficult coagulation control, and insufficient intelligent proportioning of traditional materials, achieving high-value utilization of solid waste. The material has a 3-day compressive strength ≥15MPa, a 28-day compressive strength ≥30MPa, and an apparent porosity ≤4%, possessing advantages of high strength, impermeability, and green low-carbon characteristics, suitable for the integrated needs of goaf seepage prevention and support.

Description

Technical Field

[0001] This invention belongs to the technical field of green, low-carbon, and seepage-resistant underground engineering filling materials. Specifically, it relates to the preparation and construction method of a high-strength, seepage-resistant coal-based solid waste geopolymer grouting material for filling underground mining voids, which is particularly suitable for integrated seepage prevention and support filling projects in underground mining voids. Background Technology

[0002] Underground goaf backfilling is one of the core technologies for safe mine production, and the performance of grouting materials directly determines the effectiveness of goaf remediation. Currently, the mainstream backfilling grouting materials in the industry are mainly cement-based materials and ordinary geopolymer materials. However, both types of materials have significant technical defects: cement-based grouting materials have high production costs and large carbon emissions during the production process. Moreover, in mining environments rich in acidic groundwater, calcium dissolution is prone to occur, leading to rapid attenuation of the backfill strength and poor long-term stability, which cannot meet the long-term seepage prevention and support requirements of goaf areas. Although ordinary geopolymer materials have the advantages of being green, low-carbon, and corrosion-resistant, when using coal gangue and other coal-based solid waste as raw materials, the activity of the solid waste is difficult to be effectively activated, resulting in insufficient strength of the backfill. Furthermore, the material is prone to "flash setting," and the setting time cannot be precisely controlled according to the complex underground dynamic water environment and long-distance pumping requirements, resulting in poor construction adaptability.

[0003] Meanwhile, existing material proportion optimization relies heavily on engineering experience and has a low level of intelligence. Traditional proportion prediction models based on a single backpropagation (BP) algorithm are highly sensitive to initial weights and are prone to converge to local optima, resulting in large deviations in proportion recommendations. Conventional particle swarm optimization (PSO) neural networks suffer from premature convergence and search stagnation, and lack robustness in multivariate and highly nonlinear material proportion optimization scenarios, failing to achieve dynamic and precise proportion control based on varying downhole hydrogeological conditions.

[0004] In addition, existing filling materials generally have insufficient anti-seepage performance and high porosity, which makes them prone to leakage under high water pressure. They not only fail to effectively support the surrounding rock of the goaf, but may also lead to groundwater contamination, making it difficult to achieve the dual engineering goals of "seepage prevention" and "support" in the goaf. Summary of the Invention

[0005] The purpose of this invention is to overcome the above-mentioned defects of the prior art and provide a method for preparing and constructing a high-strength, impermeable coal-based solid waste geopolymer grouting material for filling underground goaf areas.

[0006] The primary objective of this invention is to provide a coal-based solid waste geopolymer grouting material with high solid waste resource utilization rate, excellent mechanical properties, and strong impermeability, thereby solving the problems of high carbon emissions and poor corrosion resistance of traditional cement-based materials, and low solid waste activity and difficult-to-control setting time of ordinary geopolymer materials.

[0007] Another objective of this invention is to provide a preparation process for the grouting material, which fully activates the activity of coal gangue through a stepped mechanical activation process, and optimizes the microstructure of the material by combining a ternary synergistic cementing system, thereby achieving precise control of the material's setting time and synergistic improvement of its impermeability and strength properties.

[0008] Another objective of this invention is to introduce an AI-based dynamic proportioning optimization method based on the DEGI-BPNN algorithm to address the problems of existing proportioning optimization relying on experience, low intelligence, insufficient prediction accuracy and poor generalization ability of traditional neural network models. This method enables on-site dynamic and precise optimization of grouting material proportions, adapting to complex and variable hydrogeological conditions underground, and ultimately providing a green and efficient filling solution that integrates seepage prevention and support for underground goaf areas.

[0009] This invention provides a method for preparing and constructing a high-strength, impermeable coal-based solid waste polymer grouting material for filling underground goaf areas, comprising the following three steps:

[0010] Step 1, Material Composition Design: Design a grouting material consisting of component A and component B. Component A is a solid powder, and component B is a liquid activator. Component A contains five raw materials in a predetermined mass percentage: mechanically activated coal gangue powder, finely ground blast furnace slag powder, high-calcium fly ash, nano-modifier, and hydrophobic modifier. Component B contains four raw materials in predetermined mass percentages: water glass, sodium hydroxide, composite retarder, and water.

[0011] Step 2, core preparation process: Mechanically activated coal gangue powder is prepared by a step-by-step mechanical activation process of dry grinding followed by wet grinding, and then component A is obtained by homogenization and mixing in a ternary system. Component B is prepared by compound activator.

[0012] Step 3, AI dynamic ratio optimization: Based on the differential evolution-particle swarm optimization-backpropagation hybrid optimization algorithm of global information, a lightweight AI neural network model is constructed. The mixing ratio of components A and B is optimized by combining the hydrogeological conditions of the underground goaf and the grouting distance. After mixing, the mixture is grouted and filled.

[0013] Preferably, in step 1, component A, by mass percentage, consists of: 40%-60% mechanically activated coal gangue powder, 20%-30% finely ground blast furnace slag powder, 10%-20% high-calcium fly ash, 1%-5% nano-modifier, and 0.5%-2% hydrophobic modifier; component B, by mass parts, consists of: 60-80 parts water glass, 5-15 parts sodium hydroxide, 1-5 parts composite retarder, and 20-40 parts water.

[0014] Preferably, the specific method of the step-by-step mechanical activation process in step 2 is as follows: after crushing the coal gangue, dry grinding it to a specific surface area of ​​300-400 m² / kg, then mixing it with nano-modifier, process water, and grinding aid, wet grinding it to the micron or submicron level, and then drying it to obtain mechanically activated coal gangue powder.

[0015] Preferably, the preparation method of component A in step 2 is as follows: mechanically activated coal gangue powder, finely ground blast furnace slag powder, high-calcium fly ash, and hydrophobic modifier are added to a mixer according to a preset mass percentage and homogenized until there is no particle agglomeration; the preparation method of component B is as follows: sodium hydroxide is dissolved in water and cooled, then water glass and composite retarder are added and stirred until uniform and free of precipitation.

[0016] Preferably, step 3 further includes a dataset construction and preprocessing method based on a hybrid optimization algorithm of differential evolution-particle swarm optimization-backpropagation using global information, specifically:

[0017] A dataset with a total sample size of ≥2000 groups containing experimental data, literature data, and simulation data was constructed. The input features were material ratios and environmental parameters, and the output indicators were the core performance indicators of the materials. The minimum-maximum normalization method was used to map the data to the interval [0, 1].

[0018] Preferably, the method for constructing the lightweight AI neural network model in step 3 is as follows: Construct a single-hidden-layer feedforward network with 13 input neurons, 27 hidden neurons, and 6 output neurons. The hidden layer uses the Sigmoid activation function, and the output layer uses a linear activation function. The total number of parameters is 600. The number of hidden layer neurons is determined by the formula... Calculate, where, This represents the number of neurons in the hidden layer. This represents the number of neurons in the input layer.

[0019] Preferably, the differential evolution mutation enhancement method based on the global information-particle swarm optimization-backpropagation hybrid optimization algorithm in step 3 is as follows: Select the particles with the best, second-best, and third-best fitness from the particle set at iteration gen, where gen is the iteration number, according to the formula... Mutant individuals are generated, and the particle positions are constrained within the interval [0, 1] using the boundary reflection method. For the gen generation, the th A mutated individual particle For particle indexing, This is the particle with the best fitness in the gen-th generation population. The particle with the second-best fitness in the gen-th generation population. The particle with the third best fitness in the gen generation population. This is the scaling factor for the differential evolution algorithm.

[0020] Preferably, the particle swarm optimization method based on the differential evolution-particle swarm optimization-backpropagation hybrid optimization algorithm in step 3 is as follows: setting the maximum number of iterations to 300, the population size to 50, and the inertia weight to 0.3-0.9, and following the linear decreasing formula... The inertia weights are updated iteratively using individual and global extrema to update particle velocity and position. The inertia weight for the current iteration. This represents the maximum value of the inertia weight. This represents the minimum value of the inertia weight. This represents the current iteration of the algorithm. The maximum number of iterations for particle swarm optimization.

[0021] Preferably, the backpropagation local fine-tuning method of the differential evolution-particle swarm optimization-backpropagation hybrid optimization algorithm based on global information in step 3 is as follows: setting the learning rate to 0.01 and the minimum convergence error to 10. -2 According to the backpropagation direction of the error, update the weights of the hidden layer-output layer and the input layer-hidden layer and the hidden layer bias term respectively until the convergence condition is met or an overfitting trend appears.

[0022] Preferably, step 3 also includes on-site dynamic mixing and construction. The specific method is as follows: collect the ambient temperature, groundwater pH value, water flow velocity and target performance indicators in the well, input the trained AI lightweight neural network model constructed by the differential evolution-particle swarm optimization-backpropagation hybrid optimization algorithm based on global information to obtain the optimal mixing ratio of components A and B, and grout after mixing and stirring to ensure that the compressive strength of the material is ≥15MPa after 3 days of curing, ≥30MPa after 28 days of curing, and the apparent porosity is ≤4%.

[0023] The high-strength, impermeable coal-based solid waste polymer grouting material and its preparation process for filling underground goaf areas of the present invention have the following advantages compared with the prior art:

[0024] (1) This invention uses coal gangue, fly ash and other coal-based solid waste as the main raw materials, accounting for more than 70% of the total, realizing large-scale and high-value-added resource utilization of industrial solid waste, and significantly reducing the production cost of grouting materials; replacing traditional cement-based materials, significantly reducing carbon emissions in the production process, and the three-dimensional network structure of geopolymer can effectively solidify trace heavy metal ions in coal gangue, avoiding secondary environmental pollution caused by solid waste utilization, which meets the development requirements of green mine construction.

[0025] (2) By constructing a ternary synergistic cementing system of "coal gangue-slag-fly ash", the latent hydraulic properties of slag are fully utilized in the alkaline activation environment. In conjunction with the micro-aggregate effect of highly active coal gangue powder and fly ash, a large amount of dense (N,C)-ASH gel can be generated in the early stage of the material. The 3d compressive strength is ≥15MPa and the 28d compressive strength is ≥30MPa. It can quickly form effective support for the surrounding rock of the goaf and greatly reduce the safety risks such as collapse during the construction of the goaf in the mine.

[0026] (3) Through the regulation effect of composite retarder, combined with the dynamic ratio optimization technology of DEGI-BPNN algorithm, the material setting time can be precisely controlled within 30 minutes to 6 hours, which fundamentally solves the industry problem of "flash setting" of traditional geopolymer materials. It can adapt to the construction needs of different dynamic water environment and different grouting distance in the well, and meet the grouting construction requirements of long-distance pumping and special formations.

[0027] (4) The step-by-step mechanical activation process of "dry grinding followed by wet grinding" is adopted to fully destroy the layered aluminosilicate lattice structure of coal gangue and greatly improve its reactivity. Combined with the nucleus filling effect of nano-modifier and the capillary blocking effect of hydrophobic modifier, the pore structure of the solidified body is significantly optimized. The apparent porosity of the material can be controlled below 4%, and the permeability coefficient is reduced by two orders of magnitude compared with ordinary geopolymer materials. It can effectively resist acidic groundwater erosion and avoid strength decay caused by calcium dissolution, thus realizing the dual engineering functions of "seepage prevention" and "long-term stable support" in the goaf area.

[0028] (5) The DEGI-BPNN hybrid optimization algorithm proposed in this invention fundamentally solves the defects of traditional BP algorithm being prone to getting trapped in local optima and conventional PSO algorithm being prematurely converged through a three-stage training mechanism of "differential evolution mutation enhancement - particle swarm global optimization - backpropagation local fine-tuning". The model's average absolute percentage error (MAPE) for the prediction of the six core performance indicators of materials is ≤2%, and its generalization ability and robustness are significantly improved. The model is a lightweight neural network with a total of about 600 parameters. The number of hidden layer neurons is calculated by H=2n+1. On a mainstream industrial-grade industrial control computer, the calculation time for a single ratio optimization is less than 0.5 seconds. It can output the optimal ratio scheme in real time according to the measured environmental parameters and engineering target performance in the well, realize the dynamic and precise control of the grouting material ratio, and adapt to the needs of underground goaf treatment under different hydrogeological conditions.

[0029] (6) The grouting material preparation process of the present invention is simple. The steps of cascade mechanical activation, homogenization mixing, and activator preparation all adopt conventional mining engineering equipment, without the need for special high-end equipment. The process is easy to operate and control. Moreover, the raw materials are widely available and inexpensive, enabling large-scale industrial production. It has good engineering application prospects and economic and social benefits. Attached Figure Description

[0030] Figure 1 This is an overall flowchart of the preparation and construction method of a high-strength, impermeable coal-based solid waste polymer grouting material for filling underground goaf areas according to the present invention.

[0031] Figure 2 This is a flowchart of step three of the preparation and construction method of a high-strength, impermeable coal-based solid waste polymer grouting material for filling underground goaf areas according to the present invention.

[0032] Figure 3 This is a comparison chart of the MAPE results of various algorithms for the preparation and construction method of a high-strength, impermeable coal-based solid waste geopolymer grouting material for filling underground goaf areas according to the present invention. Detailed Implementation

[0033] The following detailed implementation of the high-strength, impermeable coal-based solid waste geopolymer grouting material for filling underground goaf areas and its preparation process are described in detail with reference to specific embodiments. The embodiments of the present invention are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0034] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a high-strength, impermeable coal-based solid waste polymer grouting material for filling underground goaf areas, as well as its preparation process. This invention aims to solve the core technical problems existing in the prior art:

[0035] 1. Solve the problems of high cost, large carbon emissions and poor corrosion resistance of traditional cement-based materials: Traditional cement grouting materials not only have high production costs and large carbon emissions during the production process, but also are prone to calcium dissolution in acidic groundwater environments, resulting in a decrease in the strength of the filling body and poor long-term stability.

[0036] 2. Solve the problems of difficulty in activating the activity of solid wastes such as coal gangue and difficulty in controlling the setting time of ordinary geopolymers: Direct use of coal gangue as aggregate results in low activity and insufficient strength of the filling body; while ordinary geopolymer materials are prone to "flash setting", and the setting time is difficult to accurately control according to the complex dynamic water environment and long-distance transportation requirements underground.

[0037] 3. Solve the problem of insufficient impermeability of existing backfill materials: Conventional backfill materials have high porosity and poor impermeability. They are prone to leakage under high water pressure. Not only can they not effectively support the goaf, but they may also open up waterways and pollute groundwater, failing to achieve the dual functions of "seepage prevention" and "support".

[0038] 4. Addressing the problems of existing grouting material proportion optimization relying on experience and having low intelligence, and traditional neural network models being prone to getting trapped in local optima, insufficient prediction accuracy, and poor generalization ability: Existing proportion prediction models based on a single backpropagation (BP) algorithm are highly sensitive to initial weights and are prone to converge to local minima, resulting in large deviations in proportion recommendations and failing to simultaneously meet the synergistic optimization requirements of multiple indicators such as setting time, strength, and impermeability; Conventional particle swarm optimization (PSO) neural networks suffer from premature convergence and search stagnation problems, and lack robustness in multivariate and highly nonlinear proportion optimization scenarios, making them unable to adapt to complex and variable hydrogeological conditions downhole to achieve dynamic and precise control.

[0039] To address the aforementioned technical problems, this invention provides a high-strength, impermeable coal-based solid waste geopolymer grouting material for filling underground goaf areas and its preparation process. This material achieves comprehensive performance enhancement and intelligent, precise control by constructing a ternary synergistic cementitious system, combining a "mechanical-chemical" dual activation process, and introducing an artificial intelligence (AI) dynamic proportioning optimization method based on a differential evolution-particle swarm optimization-backpropagation hybrid optimization algorithm (DEGI-BPNN) with global information.

[0040] Example 1: Implementation of a preparation and construction method for a high-strength, impermeable coal-based solid waste geopolymer grouting material for filling underground goaf areas.

[0041] like Figure 1 The diagram shown is an overall flow chart of the preparation and construction method of a high-strength, impermeable coal-based solid waste polymer grouting material for filling underground goaf areas according to the present invention. The specific scheme of the method is as follows:

[0042] Step 1, Material Composition Design.

[0043] The high-strength, impermeable coal-based solid waste polymer grouting material for filling underground goaf areas described in this invention is composed of component A (solid powder) and component B (liquid activator).

[0044] Step 1 designs a grouting material consisting of component A and component B. Component A is a solid powder, and component B is a liquid activator. Component A contains five raw materials in a predetermined mass percentage: mechanically activated coal gangue powder, finely ground blast furnace slag powder, high-calcium fly ash, nano-modifier, and hydrophobic modifier. Component B contains four raw materials in predetermined mass percentages: water glass, sodium hydroxide, composite retarder, and water.

[0045] In step 1, component A, by mass percentage, consists of: 40%-60% mechanically activated coal gangue powder, 20%-30% finely ground blast furnace slag powder, 10%-20% high-calcium fly ash, 1%-5% nano-modifier, and 0.5%-2% hydrophobic modifier; component B, by mass parts, consists of: 60-80 parts water glass, 5-15 parts sodium hydroxide, 1-5 parts composite retarder, and 20-40 parts water. The specific process is as follows:

[0046] Step 11, Component A preparation.

[0047] The A component (solid powder) includes, by mass percentage: (1) mechanically activated coal gangue powder: 40%-60%; (2) finely ground blast furnace slag powder: 20%-30%; (3) high-calcium fly ash: 10%-20%; (4) nano-modifier: 1%-5%; (5) hydrophobic modifier: 0.5%-2%.

[0048] Step 12, Component B preparation.

[0049] The B component (liquid activator) includes, by mass, the following components: (1) water glass (modulus 1.2-1.5): 60-80 parts; (2) sodium hydroxide: 5-15 parts; (3) composite retarder: 1-5 parts; (4) water: 20-40 parts.

[0050] The present invention has a high solid waste resource utilization rate and significant green and low-carbon benefits: The present invention uses coal gangue, fly ash and other coal-based solid wastes as the main raw materials (accounting for more than 70% in total), realizing large-scale, high-value-added resource utilization of industrial solid waste, significantly reducing material production costs, and reducing carbon emissions and environmental pollution caused by cement production; at the same time, the geopolymer three-dimensional network structure can effectively solidify trace heavy metal ions in coal gangue, avoiding secondary pollution.

[0051] Step 2, core preparation process steps.

[0052] Step 2 involves preparing mechanically activated coal gangue powder through a stepped mechanical activation process of first dry grinding and then wet grinding. Component A is then obtained by homogenization and mixing in a ternary system, and component B is prepared by compound activator formulation.

[0053] The specific method of the cascade mechanical activation process in step 2 is as follows: after crushing the coal gangue, dry grinding is carried out to a specific surface area of ​​300-400 m² / kg, and then mixed with nano-modifier, process water and grinding aid for wet ultrafine grinding to micron or submicron level. After drying, mechanically activated coal gangue powder is obtained.

[0054] In step 2, component A is prepared by adding mechanically activated coal gangue powder, finely ground blast furnace slag powder, high-calcium fly ash, and hydrophobic modifier to a mixer according to a preset mass percentage, and homogenizing until no particle agglomeration occurs. Component B is prepared by dissolving sodium hydroxide in water, cooling it, adding water glass and a composite retarder, and stirring until homogeneous and free of precipitation. The specific process is as follows:

[0055] Step 21, cascade mechanical activation pretreatment.

[0056] The activation treatment of coal gangue is carried out using a "dry grinding followed by wet grinding" process, specifically including:

[0057] (1) Dry grinding: After crushing the coal gangue, it is put into a ball mill for the first dry grinding, so that its specific surface area reaches 300-400 m² / kg, and its stable layered aluminosilicate lattice structure is initially destroyed to obtain dry-ground coal gangue powder.

[0058] (2) Wet grinding: The above-mentioned dry-ground coal gangue powder, nano-modifier, appropriate amount of process water and grinding aid are mixed and placed in a stirred mill for a second wet ultrafine grinding. By using mechanochemical action, the particles are further refined to the micron level or even the submicron level. At the same time, a large number of active sites are introduced on the surface of the new particles, which significantly reduces the activation energy of the subsequent polymerization reaction, and mechanically-activated coal gangue slurry is obtained. After drying, mechanically activated coal gangue powder is obtained.

[0059] Step 22: Homogenize and mix the ternary system.

[0060] According to the mass percentage of step 11, the mechanically activated coal gangue powder, finely ground blast furnace slag powder, high-calcium fly ash, and hydrophobic modifier obtained in step 21 are added into the mixer and mixed thoroughly to obtain homogeneous component A (solid powder).

[0061] Step 23, preparation of the composite activator solution.

[0062] According to the mass fraction of step 12, dissolve sodium hydroxide in water, cool it, add water glass and composite retarder, stir evenly to obtain component B (liquid activator), wherein the composite retarder is used to precisely control the gelation time and prevent "flash coagulation".

[0063] Step 24, Dynamic mix design and grouting construction.

[0064] Based on the specific hydrogeological conditions (water temperature, pH value, water flow velocity) and grouting distance of the underground goaf, a pre-trained AI lightweight neural network model based on the DEGI-BPNN algorithm is used to optimize and recommend the mixing ratio of component A and component B in real time. During construction, component A and component B are mixed according to the dynamic ratio recommended by the AI ​​model, and after being stirred evenly by a high-speed mixer, grouting is carried out.

[0065] The "step-by-step mechanical activation" process significantly enhances the reactivity of coal gangue. Combined with the nucleus-filling effect of nano-modifiers and the capillary blocking effect of hydrophobic modifiers, the pore structure of the solidified body is significantly optimized. The apparent porosity of the solidified material can be controlled below 4%, and the permeability coefficient is reduced by two orders of magnitude compared with ordinary geopolymer materials. It can effectively resist acidic groundwater erosion and avoid strength attenuation caused by calcium dissolution, thus achieving the dual functions of "seepage prevention" and "long-term stable support".

[0066] By constructing a ternary synergistic cementing system of "coal gangue (skeleton) - slag (early strength) - fly ash (filler)", combined with a composite retarder and the DEGI-BPNN algorithm dynamic proportioning optimization technology proposed in this invention, the industry problem of "flash setting" of traditional geopolymer materials is fundamentally solved. The material setting time can be precisely controlled within 30 minutes to 6 hours according to the actual working conditions such as the underground dynamic water environment and grouting distance, so as to meet the grouting construction needs of long-distance pumping and special strata.

[0067] Step 3, an AI dynamic allocation optimization method based on the differential evolution-particle swarm optimization-backpropagation hybrid optimization algorithm (DEGI-BPNN) with global information. For example... Figure 2 The diagram shows a flowchart of step three of the preparation and construction method of a high-strength, impermeable coal-based solid waste geopolymer grouting material for filling underground goaf areas according to the present invention.

[0068] This invention proposes a global information differential evolution-particle swarm optimization-backpropagation hybrid optimization algorithm (DEGI-BPNN), and builds an AI lightweight neural network model based on this algorithm to achieve real-time dynamic optimization of grouting material ratio. Specifically, it includes four core steps: dataset construction and preprocessing, neural network model structure design, DEGI-BPNN hybrid optimization training, and on-site dynamic ratio optimization.

[0069] Step 3 is based on a differential evolution-particle swarm optimization-backpropagation hybrid optimization algorithm with global information to construct a lightweight AI neural network model. The mixing ratio of components A and B is optimized by combining the hydrogeological conditions of the underground goaf with the grouting distance, and then the mixture is grouted and filled.

[0070] Step 3 also includes a dataset construction and preprocessing method based on the differential evolution-particle swarm optimization-backpropagation hybrid optimization algorithm with global information. Specifically, a dataset with a total sample size of ≥2000 groups containing experimental data, literature data, and simulation data is constructed. The input features are material ratios and environmental parameters, and the output indicators are the core performance indicators of the materials. The minimum-maximum normalization method is used to map the dataset to the [0, 1] interval.

[0071] In step 3, the lightweight AI neural network model is constructed as follows: a single-hidden-layer feedforward network with 13 input neurons, 27 hidden neurons, and 6 output neurons is constructed. The hidden layer uses the Sigmoid activation function, and the output layer uses a linear activation function, with a total of 600 parameters. The number of hidden layer neurons is determined by the formula... Calculate, where, This represents the number of neurons in the hidden layer. This represents the number of neurons in the input layer.

[0072] In step 3, the differential evolution mutation enhancement method based on the global information-particle swarm optimization-backpropagation hybrid optimization algorithm is as follows: Select the best, second best, and third best particles from the particle set at iteration gen, where gen is the iteration number, according to the formula... Mutant individuals are generated, and the particle positions are constrained within the interval [0, 1] using the boundary reflection method. For the gen generation, the th A mutated individual particle For particle indexing, This is the particle with the best fitness in the gen-th generation population. The particle with the second-best fitness in the gen-th generation population. The particle with the third best fitness in the gen generation population. This is the scaling factor for the differential evolution algorithm.

[0073] In step 3, the particle swarm optimization method based on the differential evolution-particle swarm optimization-backpropagation hybrid optimization algorithm with global information is as follows: The maximum number of iterations is set to 300, the population size to 50, and the inertia weight to 0.3-0.9, following a linear decreasing formula. The inertia weights are updated iteratively using individual and global extrema to update particle velocity and position. The inertia weight for the current iteration. This represents the maximum value of the inertia weight. This represents the minimum value of the inertia weight. This represents the current iteration of the algorithm. The maximum number of iterations for particle swarm optimization.

[0074] In step 3, the backpropagation local fine-tuning method of the differential evolution-particle swarm optimization-backpropagation hybrid optimization algorithm based on global information is as follows: set the learning rate to 0.01 and the minimum convergence error to 10. -2 According to the backpropagation direction of the error, update the weights of the hidden layer-output layer and the input layer-hidden layer and the hidden layer bias term respectively until the convergence condition is met or an overfitting trend appears.

[0075] Step 3 also includes on-site dynamic mixing and construction. The specific method is as follows: Collect the ambient temperature, groundwater pH, water flow velocity, and target performance indicators from the well. Input these data into a trained AI lightweight neural network model constructed using a differential evolution-particle swarm optimization-backpropagation hybrid optimization algorithm based on global information to obtain the optimal mixing ratio of components A and B. After mixing and stirring, grouting is performed to ensure that the material's compressive strength after 3 days of curing is ≥15MPa, its compressive strength after 28 days of curing is ≥30MPa, and its apparent porosity is ≤4%. The specific process is as follows:

[0076] Step 31, Dataset Construction and Preprocessing.

[0077] Step 311, Data Source.

[0078] The dataset consists of three parts: experimental data, literature data, and simulation data. To ensure the diversity, coverage, and authenticity of the samples, the total number of valid data samples is ≥2000. The data acquisition methods for each part are as follows: (1) Experimental data: obtained through indoor orthogonal experiments and single-factor variable experiments, covering the key variable range of material ratio and environmental parameters. A total of about 500 valid data samples were obtained to provide core real sample support for the dataset; (2) Literature data: data on the ratio and performance of polymer grouting materials for coal-based solid waste bases were collected from publicly available academic literature. After outlier removal, dimension unification, and consistency verification, about 300 valid samples were expanded to improve the scene coverage of the samples; (3) Simulation data: a Box-Behnken response surface model was constructed based on the previous experimental data. Latin hypercube uniform space filling sampling was carried out within a reasonable range of input parameters to generate about 1200 high-confidence simulation samples, solving the problem of insufficient experimental sample size under multivariate combination and enhancing the stability and generalization ability of model training.

[0079] In this step, the input variables are 13 features, including the proportion of mechanically activated coal gangue powder, finely ground blast furnace slag powder, high-calcium fly ash, nano-modifier, and hydrophobic modifier in component A; the water glass modulus, water glass content, sodium hydroxide content, composite retarder content, and water-cement ratio in component B; as well as the ambient temperature, groundwater pH value, and water flow velocity. The output indicators are 6 core performance indicators, including material setting time, fluidity, 3-day compressive strength, 28-day compressive strength, apparent porosity, and permeability coefficient.

[0080] Step 312, data preprocessing.

[0081] The min-max normalization method is used to linearly map all input and output feature data to... To eliminate the influence of different units of measurement on model training, the normalization formula is as follows:

[0082] ;

[0083] In the formula, : Normalized eigenvalues These are the original eigenvalues. The minimum value of this feature in the dataset. The maximum value of this feature in the dataset. , These are the upper and lower limits of the target interval after normalization.

[0084] Step 32, Lightweight Neural Network Model Structure Design.

[0085] A multi-layer feedforward neural network with a single hidden layer is constructed to meet the lightweight deployment requirements of downhole field control computers / embedded systems. The model structure parameters and calculation rules are defined as follows:

[0086] Step 321, Input layer design.

[0087] The number of neurons in the input layer is denoted as , Each of these corresponds one-to-one with the 13 input feature dimensions in step 311.

[0088] Step 322, hidden layer design.

[0089] The number of neurons in the hidden layer is denoted as , Using empirical formulas Calculation, i.e. The activation function for hidden layer neurons is the Sigmoid function. The formula for calculating the output of each hidden layer neuron is:

[0090] ;

[0091] In the formula, For the input layer The first neuron to the hidden layer The connection weights of each neuron For the input layer The input values ​​of each feature, For the hidden layer The bias terms for each neuron are defined using only the symbols in this formula and do not refer to other parameters repeatedly.

[0092] The number of neurons in the input layer, with a value of 13, corresponds one-to-one with the 13 input feature dimensions; The number of neurons in the hidden layer is given by an empirical formula. The calculation yields 27; : The index of the hidden layer neuron, with values ​​1, 2, ... ; : The index of the input layer neuron, with values ​​1, 2, ... ; The number of neurons in the output layer, with a value of 6, corresponds one-to-one with the 6 core performance output indicators; : The index of the output layer neuron, with values ​​of 1, 2, ... ; : No. The output values ​​of each hidden layer neuron; : Natural exponential function.

[0093] Step 323, Output layer design.

[0094] The number of neurons in the output layer is denoted as , This corresponds one-to-one with the six core performance output metrics in step 311; the output layer activation function uses a linear activation function to adapt to the requirements of regression prediction tasks. The formula for calculating the output of each neuron in the output layer is:

[0095] , ;

[0096] In the formula, For the hidden layer The first neuron to the output layer The connection weights of each neuron For the hidden layer defined in step 322 Each neuron outputs a value, and each symbol is used only for the definition of this formula and does not refer to other parameters repeatedly. : No. The output values ​​of each output layer neuron; Hidden layer The first neuron to the output layer The connection weights of each neuron.

[0097] Step 324, design the loss function and fitness function.

[0098] Mean squared error (MSE) is used as the model loss function to measure the deviation between the model's predicted values ​​and the measured values. The calculation formula is as follows:

[0099] ;

[0100] In the formula, For the first The model's predicted output value for each sample. For the first The measured target value for each sample. For the first The prediction error of each sample; the fitness function for model training is defined as Fitness. In the formula For a single particle in a particle swarm, the smaller the fitness value, the higher the model's prediction accuracy. : Index of samples; Fitness : Particle fitness value; MSE: Mean Squared Error, which measures the deviation between model predictions and actual values.

[0101] Step 325, Lightweight Verification.

[0102] The neural network constructed in this invention has a total of approximately 600 parameters, a compact model structure, and low computational complexity. Tests have shown that on mainstream industrial-grade control computers (such as Intel Celeron 2.0GHz processors), the calculation time for a single ratio optimization is less than 0.5 seconds, fully meeting the lightweight deployment requirements for real-time dynamic optimization on-site.

[0103] Step 33, DEGI-BPNN hybrid optimization algorithm training process.

[0104] The DEGI-BPNN algorithm proposed in this invention optimizes the weights and biases of the neural network through a three-stage training process: differential evolution mutation enhancement, particle swarm optimization, and backpropagation local fine-tuning. This overcomes the shortcomings of traditional BP algorithms, which are prone to getting trapped in local optima, and PSO algorithms, which suffer from premature convergence. The specific training steps are as follows:

[0105] Step 331, algorithm parameter initialization.

[0106] Initialize the core parameters for neural network training and the DEGI-BPNN algorithm. The values ​​and definitions of each parameter are as follows:

[0107] (1) BP algorithm parameters: The learning rate is denoted as , The minimum convergence error is denoted as , ;

[0108] (2) Particle Swarm Optimization (PSO) Algorithm Parameters: The maximum number of iterations is denoted as... , Population size is denoted as , The maximum value of the inertia weight is denoted as The minimum value is denoted as , , Cognitive learning factors are denoted as Social learning factors are denoted as , , This value is based on the optimal parameter combination that has been experimentally verified in the original research of the DEGI-BPNN algorithm; the maximum speed is denoted as... , ;

[0109] (3) Differential Evolution (DE) Algorithm Parameters: Scaling factor denoted as , This value ensures that the mutation operator has a sufficiently strong perturbation capability to effectively enhance population diversity, while avoiding excessive mutation that could lead to algorithm degradation.

[0110] All weights of the neural network (input layer - hidden layer weights) Hidden layer - output layer weights The bias term is mapped to the position vector of a particle in the PSO algorithm, and the dimension of a single particle is denoted as . , The initial position and velocity of all particles are randomly generated within the interval [0, 1].

[0111] in, The learning rate for the BP algorithm is 0.01. The minimum convergence error of the BP algorithm is denoted as: ; : Maximum number of iterations for the PSO algorithm, with a value of 300; NP: Population size for the PSO algorithm, with a value of 50; The maximum value of the inertia weight in the PSO algorithm is 0.9. The minimum value of the inertia weight in the PSO algorithm is 0.3. The cognitive learning factor of the PSO algorithm is set to 1.56. The social learning factor of the PSO algorithm is 1.56. : The maximum speed of the PSO algorithm, with a value of 1; The scaling factor for the DE algorithm is 0.5.

[0112] Step 332, differential evolution enhancement operation based on improved mutation operator.

[0113] An improved differential evolution mutation operator is introduced to perturb the particle positions, enhancing population diversity and overcoming the premature convergence problem of the PSO algorithm, which is prone to getting trapped in local optima in the later stages of iteration. The specific execution steps are as follows:

[0114] Step 3321: Optimal individual selection.

[0115] In the gen generation population, based on the fitness value of each particle... Sort the particles in ascending order (smaller fitness values ​​indicate better particle performance), and select the particle with the best fitness (i.e., the smallest prediction error). The suboptimal particle is denoted as The third best particle is denoted as .

[0116] Step 3322, differential mutation operation.

[0117] Differential mutation is performed on the three optimal individuals selected in step 3321 to generate a corresponding mutated individual for each particle. The mutation formula is as follows:

[0118] ;

[0119] In the formula, The scaling factor for the DE algorithm defined in step 331 is used by the mutation operator to guide the search by using information about the best individual in the population. This not only preserves the good genes of the best individual, but also introduces a new search direction through the difference vector, which effectively improves the global exploration capability of the algorithm and solves the search stagnation problem in the PSO iteration process.

[0120] Where gen: the iteration number of the algorithm; The particle with the best fitness in the gen-th generation population; The particle with the second-best fitness in the gen-th generation population; The particle with the third best fitness in the gen generation population; : in the gen generation A mutated individual of a particle.

[0121] Step 3323, Boundary condition constraints.

[0122] Boundary condition checks are performed on the mutated individuals to ensure that their positions are within the predefined search space. Inside, among which , If a mutated individual exceeds the boundary, the boundary reflection method is used to pull it back into the feasible region, as shown in the following formula:

[0123] ;

[0124] In the formula, , The optimal particle is consistent with the one defined in step 3321; 1.8 is the reflection coefficient, which is determined through pre-experimentation to ensure population diversity while avoiding excessive particle bounce and ensuring that the particle always completes the optimization within the legal search space.

[0125] in, : The upper limit of the particle position, with a value of 1; : The lower limit of the particle's position, with a value of 0; The maximum reference value in the boundary reflection method is [value]. ; The minimum reference term in the boundary reflection method takes the value of .

[0126] Step 333: Particle swarm optimization iteration.

[0127] Based on the mutated population from step 332, perform PSO iterative optimization to search for the globally optimal approximate solution of the neural network weights and biases in the solution space. The specific steps are as follows:

[0128] Step 3331: Update individual and global extrema.

[0129] For each particle, calculate its fitness value and update its historical best position. (Local optimum) and global optimum position of the population (Globally optimal), the update formula is as follows:

[0130] ;

[0131] In the formula, The population size defined in step 331; : No. The individual historical best position of each particle; The global optimal position of the population.

[0132] Step 3332, Particle velocity and position update.

[0133] The algorithm employs a linearly decreasing inertia weight balancing mechanism, which combines global exploration and local exploitation capabilities. The inertia weight calculation formula is as follows:

[0134] ;

[0135] In the formula, , , These are the PSO algorithm parameters defined in step 331;

[0136] Based on the updated inertia weights, the particle velocity and position are iteratively updated, as shown in the following formula:

[0137] ;

[0138] In the formula, The velocity of the particle after the update. The updated position of the particle. , The learning factor for the PSO algorithm defined in step 331. , for A random number that is uniformly distributed within an interval.

[0139] in, Inertia weights in the PSO algorithm; : in the gen generation The velocity of each particle; : in the gen generation The position of each particle.

[0140] Step 3333: Iteration termination check.

[0141] If the number of iterations reaches Or the global optimal fitness value of the population is less than the minimum convergence error defined in step 331. Terminate the PSO iteration and output the globally optimal position. , as the initial optimal seed for the neural network weights and biases; otherwise, return to step 332 to continue performing the differential evolution enhancement operation.

[0142] Step 334, local fine-tuning based on backpropagation of gradient descent.

[0143] Using the globally optimal weights and biases output in step 333 as the initial values ​​for the backpropagation (BP) algorithm, gradient descent backpropagation is performed for local fine-tuning to further reduce model prediction error. The specific steps are as follows:

[0144] Step 3341, error backpropagation and parameter update.

[0145] Based on the prediction error output by the model, the weights and bias terms of the hidden layer-output layer and input layer-hidden layer are updated in the direction of backpropagation of the error, respectively. The update rules for each parameter are as follows:

[0146] (1) Connection weights between hidden layer and output layer :

[0147] ;

[0148] In the formula, The learning rate for the BP algorithm is defined in step 331. For the index of output layer neurons, For indexing hidden layer neurons, For the output layer The predicted value of each neuron, To correspond to the measured target value, For the hidden layer defined in step 322 The output value of each neuron For weight update amount, The weight values ​​before the update. The updated weight values;

[0149] (2) Connection weights between the hidden layer and the input layer :

[0150] ;

[0151] In the formula, For the index of input layer neurons, The derivative of the sigmoid activation function in the hidden layer. For the input layer The input values ​​of each feature, For weight update amount, The weight values ​​before the update. The updated weight values ​​are used, and the remaining symbols are consistent with those defined in step (1) of this procedure.

[0152] (3) Hidden layer bias terms :

[0153] ;

[0154] In the formula, For the amount of bias term update, The bias value before the update. The value is the updated bias term, and the other symbols are consistent with the definitions in steps (1) and (2) of this procedure.

[0155] Step 3342, fine-tune the termination judgment.

[0156] If the global fitness value of the model is less than that defined in step 331 If an overfitting trend occurs during the iteration process (the validation set error increases for 5 consecutive rounds), terminate the BP fine-tuning to obtain the trained DEGI-BPNN lightweight neural network model; otherwise, repeat step 3341 to continue parameter updates.

[0157] Step 335, Model Cross-Validation. (Combined with Appendix) Figure 3 As shown, Figure 3 This is a comparison chart of the MAPE results of various algorithms for the preparation and construction method of a high-strength, impermeable coal-based solid waste geopolymer grouting material for filling underground goaf areas according to the present invention.

[0158] A 10-fold cross-validation method was used to verify the performance of the trained DEGI-BPNN model. The dataset was randomly divided into 10 subsets, and 9 subsets were used as the training set and 1 subset as the test set in turn. The experiment was repeated 5 times independently, and the average result was used as the final performance index of the model to ensure the generalization ability and robustness of the model. To verify the performance superiority of the DEGI-BPNN algorithm of this invention, backpropagation neural network (BPNN) and particle swarm optimization backpropagation neural network (PSO-BPNN) were introduced as baseline comparison algorithms. Comparative experiments were carried out on the same dataset, with the same training parameters and validation methods, and the mean absolute percentage error (MAPE) of each algorithm on the training set and the test set was calculated. Testing revealed that the DEGI-BPNN model constructed in this invention controlled the mean absolute percentage error (MAPE) of six output indicators, including condensation time, fluidity, compressive strength, apparent porosity, and permeability coefficient, to within 2%. Comparative analysis showed that the DEGI-BPNN algorithm proposed in this invention had significantly lower MAPE values ​​on both the training and test sets compared to the two baseline algorithms, PSO-BPNN and BPNN. Specifically, the MAPE value of the DEGI-BPNN algorithm on the test set was only 1.8%, which is 55% lower than that of the PSO-BPNN algorithm and 74.3% lower than that of the BPNN algorithm. At the same time, the DEGI-BPNN algorithm had the smallest difference in MAPE between the training and test sets, at only 0.2%, which is far lower than that of PSO-BPNN (0.8%) and BPNN (2.1%). This indicates that the present invention, through a three-stage training mechanism of differential evolution mutation enhancement, particle swarm optimization, and backpropagation local fine-tuning, effectively solves the problems of traditional BPNN algorithms being prone to getting trapped in local optima and having low prediction accuracy. It also overcomes the defects of PSO-BPNN algorithm, such as premature convergence and insufficient generalization ability. This enables the model to maintain high training accuracy while having stronger test set prediction ability and engineering generalization ability, fully meeting the engineering application requirements of dynamic optimization of downhole grouting material ratio.

[0159] Step 34: On-site dynamic proportioning optimization application.

[0160] Before the underground grouting construction, environmental parameters (ambient temperature, groundwater pH value, and water flow velocity) and target performance indicators required by the engineering design (target setting time, minimum compressive strength, maximum apparent porosity, and maximum permeability coefficient) are collected on-site. These parameters are then input into the DEGI-BPNN model trained in step 33. The model aims to minimize the deviation between the predicted performance and the target performance. Through global optimization, it quickly outputs the optimal ratio scheme of component A and component B that meets the engineering requirements, guiding the preparation and construction of grouting materials on-site.

[0161] This invention proposes a DEGI-BPNN hybrid optimization algorithm and constructs a lightweight neural network model. Through a three-stage training mechanism of "differential evolution enhancement - particle swarm optimization - BP local fine-tuning", it fundamentally solves the defects of traditional BP algorithm being prone to getting trapped in local optima and conventional PSO algorithm being prematurely convergent. The model's average absolute percentage error in predicting material properties is ≤2%, and its generalization ability and robustness are significantly improved. The model can realize real-time dynamic optimization of grouting material ratio according to the complex and variable hydrogeological conditions and engineering performance requirements in the well, providing a flexible, reliable and intelligent solution for the treatment of goaf under different geological conditions.

[0162] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0163] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for preparing and constructing a high-strength, impermeable coal-based solid waste polymer grouting material for filling underground goaf areas, characterized in that, It includes the following three steps: Step 1, Material Composition Design: Design a grouting material consisting of component A and component B. Component A is a solid powder, and component B is a liquid activator. Component A contains five raw materials in a predetermined mass percentage: mechanically activated coal gangue powder, finely ground blast furnace slag powder, high-calcium fly ash, nano-modifier, and hydrophobic modifier. Component B contains four raw materials in predetermined mass percentages: water glass, sodium hydroxide, composite retarder, and water. Step 2, core preparation process: Mechanically activated coal gangue powder is prepared by a step-by-step mechanical activation process of dry grinding followed by wet grinding, and then component A is obtained by homogenization and mixing in a ternary system. Component B is prepared by compound activator. Step 3, AI dynamic ratio optimization: Based on the differential evolution-particle swarm optimization-backpropagation hybrid optimization algorithm of global information, a lightweight AI neural network model is constructed. The mixing ratio of components A and B is optimized by combining the hydrogeological conditions of the underground goaf and the grouting distance. After mixing, the mixture is grouted and filled.

2. The preparation and construction method of the high-strength, impermeable coal-based solid waste polymer grouting material for filling underground goaf areas according to claim 1, characterized in that, In step 1, component A, by mass percentage, consists of: 40%-60% mechanically activated coal gangue powder, 20%-30% finely ground blast furnace slag powder, 10%-20% high-calcium fly ash, 1%-5% nano-modifier, and 0.5%-2% hydrophobic modifier; component B, by mass parts, consists of: 60-80 parts water glass, 5-15 parts sodium hydroxide, 1-5 parts composite retarder, and 20-40 parts water.

3. The preparation and construction method of the high-strength, impermeable coal-based solid waste polymer grouting material for filling underground goaf areas according to claim 1, characterized in that, The specific method of the cascade mechanical activation process in step 2 is as follows: after crushing the coal gangue, dry grind it to a specific surface area of ​​300-400 m² / kg, then mix it with nano-modifier, process water and grinding aid and wet grind it to micron or submicron level, and then dry it to obtain mechanically activated coal gangue powder.

4. The preparation and construction method of the high-strength, impermeable coal-based solid waste polymer grouting material for filling underground goaf areas according to claim 1, characterized in that, The preparation method of component A in step 2 is as follows: mechanically activated coal gangue powder, finely ground blast furnace slag powder, high-calcium fly ash, and hydrophobic modifier are added to a mixer according to a preset mass percentage and homogenized until there is no particle agglomeration; the preparation method of component B is as follows: sodium hydroxide is dissolved in water and cooled, then water glass and composite retarder are added and stirred until homogeneous and free of precipitation.

5. The preparation and construction method of the high-strength, impermeable coal-based solid waste polymer grouting material for filling underground goaf areas according to claim 1, characterized in that, Step 3 also includes a dataset construction and preprocessing method based on a hybrid optimization algorithm of differential evolution-particle swarm optimization-backpropagation with global information, specifically: A dataset with a total sample size of ≥2000 groups containing experimental data, literature data, and simulation data was constructed. The input features were material ratios and environmental parameters, and the output indicators were the core performance indicators of the materials. The minimum-maximum normalization method was used to map the data to the interval [0, 1].

6. The preparation and construction method of the high-strength, impermeable coal-based solid waste polymer grouting material for filling underground goaf areas according to claim 1, characterized in that, The method for constructing the lightweight AI neural network model in step 3 is as follows: Construct a single-hidden-layer feedforward network with 13 input neurons, 27 hidden neurons, and 6 output neurons. The hidden layer uses the Sigmoid activation function, and the output layer uses a linear activation function. The total number of parameters is 600. The number of hidden layer neurons is determined by the formula... Calculate, where, This represents the number of neurons in the hidden layer. This represents the number of neurons in the input layer.

7. The preparation and construction method of the high-strength, impermeable coal-based solid waste polymer grouting material for filling underground goaf areas according to claim 1, characterized in that, In step 3, the differential evolution mutation enhancement method based on the global information-particle swarm optimization-backpropagation hybrid optimization algorithm is as follows: From the particle set at iteration gen, select the particles with the best, second-best, and third-best fitness, where gen is the iteration number, according to the formula... Mutant individuals are generated, and the particle positions are constrained within the interval [0, 1] using the boundary reflection method. For the gen generation, the th A mutated individual particle For particle indexing, This is the particle with the best fitness in the gen-th generation population. The particle with the second-best fitness in the gen-th generation population. The particle with the third best fitness in the gen generation population. This is the scaling factor for the differential evolution algorithm.

8. The preparation and construction method of the high-strength, impermeable coal-based solid waste polymer grouting material for filling underground goaf areas according to claim 1, characterized in that, In step 3, the particle swarm optimization method based on the differential evolution-particle swarm optimization-backpropagation hybrid optimization algorithm with global information is as follows: Set the maximum number of iterations to 300, the population size to 50, and the inertia weight to 0.3-0.9, and apply a linearly decreasing formula. The inertia weights are updated iteratively using individual and global extrema to update particle velocity and position. The inertia weight for the current iteration. This represents the maximum value of the inertia weight. This represents the minimum value of the inertia weight. This represents the current iteration of the algorithm. The maximum number of iterations for particle swarm optimization.

9. The preparation and construction method of the high-strength, impermeable coal-based solid waste polymer grouting material for filling underground goaf areas according to claim 1, characterized in that, The back propagation local fine adjustment method of the differential evolution-particle swarm-back propagation hybrid optimization algorithm based on global information in step 3 is: setting the learning rate to 0.01, the minimum convergence error to 10 -2 , updating the weights of the hidden layer-output layer and the input layer-hidden layer and the hidden layer bias term in the error back propagation direction, respectively, until the convergence condition is met or the overfitting trend occurs.

10. The preparation and construction method of the high-strength, impermeable coal-based solid waste polymer grouting material for filling underground goaf areas according to claim 1, characterized in that, Step 3 also includes on-site dynamic mixing and construction. The specific method is as follows: collect the ambient temperature, groundwater pH value, water flow velocity and target performance indicators in the well, input the trained AI lightweight neural network model constructed by the differential evolution-particle swarm optimization-backpropagation hybrid optimization algorithm based on global information to obtain the optimal mixing ratio of components A and B, and grout after mixing and stirring to ensure that the compressive strength of the material is ≥15MPa after 3 days of curing, ≥30MPa after 28 days of curing, and the apparent porosity is ≤4%.

Images