A high-efficiency energy storage pile for cross-seasonal energy storage

Abstract

This invention discloses a high-efficiency cross-seasonal energy storage pile, relating to the fields of new energy and energy conservation technology. It solves the problems of large heat load prediction errors, groundwater heat diffusion attenuation, reduced heat exchange capacity, and insufficient geological adaptability in existing energy storage systems. This solution uses a hybrid time-series prediction model (LA) to optimize load prediction, combines a distributed fiber optic temperature measurement array with a finite element inversion algorithm to accurately reconstruct the ground temperature field, and identifies geological types based on a support vector machine classifier to achieve adaptive control of heat exchange parameters. Simultaneously, it optimizes heat pump power, circulating pump frequency, and phase change material charge / discharge rates through a deep deterministic gradient algorithm and a gradient heat release strategy. This invention significantly improves the heat exchange efficiency, long-term stability, and adaptability to complex environments of cross-seasonal energy storage systems, thereby enhancing the energy-saving effect of building energy supply systems.

Description

Technical Field

[0001] This invention relates to the field of new energy and energy-saving technology, and more specifically to an energy pile for efficient cross-seasonal energy storage. Background Technology

[0002] Cross-seasonal energy storage piles are an energy storage technology that utilizes the thermal capacity of underground soil or water bodies to store and release heat between different seasons. Combining building foundation piles with ground source heat pump technology, energy piles leverage the thermal buffering capacity of underground soil to store heat in summer and release it in winter, thereby reducing a building's dependence on external energy sources, improving energy efficiency, and reducing carbon emissions. In recent years, with advancements in materials technology, heat exchange efficiency, and intelligent control algorithms, the performance of energy pile systems has been continuously optimized, and their application scope has expanded from individual buildings to regional heating systems.

[0003] Currently, high-efficiency cross-seasonal energy storage systems primarily rely on a combination of ground source heat pumps (GSHP) and energy pipes for energy storage and release. The core technologies of this system involve the installation method of the heat exchange pipes, the selection of the heat transfer medium, the optimized design of geological thermal properties, and the regulation of the intelligent control system. Typically, the energy pipes are laid concurrently with the building foundation construction, forming an underground thermal energy storage network. In summer, waste heat from the building or solar collectors are used to store heat in the underground medium (such as sand, clay, or rock); in winter, the stored heat is extracted by the ground source heat pump for building heating.

[0004] Among existing related patent technologies, such as CN116094154A, an energy storage management system is proposed. This system continuously collects operational status data from energy storage devices and processes and judges this data to control the operational status of the energy storage devices. Another patent, CN117791687B, proposes an energy management method for photovoltaic energy storage systems. This method estimates the cumulative power generation over a certain period by constructing a model of photovoltaic power generation to compensate for the energy consumption gap of electricity users. However, these technologies have encountered some problems when applied to energy piles for cross-seasonal energy storage.

[0005] First, existing control systems suffer from significant errors in heat load prediction. Due to dynamic changes in building usage patterns and uncertainties in weather conditions, these systems struggle to accurately match heat input and output, leading to excessive heat storage in summer or insufficient heating in winter. After long-term operation, the nonlinear effect of underground heat diffusion causes energy storage degradation, and heat exchange efficiency declines year by year. Furthermore, existing intelligent control algorithms still lack adaptive optimization capabilities, making it difficult to achieve efficient energy storage management under various geological and climatic conditions. Most systems rely on fixed heat input / output strategies, ignoring dynamic changes in ground temperature, resulting in reduced heat exchange capacity and even requiring additional energy compensation, thus lowering the overall energy-saving effect of the system. Simultaneously, because energy piles depend on the thermal conductivity of the underground soil, different geological conditions have a significant impact on system performance. For example, in sandy soil or areas with high groundwater flow velocities, heat may be rapidly dissipated, reducing overall energy storage capacity; while in clay or rock strata, heat accumulation may lead to abnormally high ground temperatures, affecting system heat exchange efficiency and even causing foundation deformation problems. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention discloses a high-efficiency cross-seasonal energy storage energy pile, aiming to solve the problems mentioned in the background technology.

[0007] To achieve the above-mentioned technical effects, the present invention adopts the following technical solution:

[0008] An efficient cross-seasonal energy storage energy pile includes: a dynamic coupling prediction module, used to perform multi-scale prediction of building heat load through an LA hybrid time series prediction model; when the building energy consumption monitoring system detects that the start-stop frequency of indoor temperature control equipment exceeds the threshold, the LA hybrid time series prediction model reconstructs the weight of historical load samples through a sliding time window mechanism, and simultaneously inputs meteorological forecast data into a three-dimensional heat conduction correction coefficient matrix to output cross-seasonal heat storage demand prediction values.

[0009] The formation temperature field reconstruction module is used to collect formation temperature field gradient data and axial and radial temperature gradients within the energy pile using a distributed fiber optic temperature measurement array. When the rate of change of the formation temperature gradient exceeds a preset critical value, the module reconstructs the three-dimensional unsteady temperature field based on the heat flux density sensor and temperature sensor data within the energy pile, calculates the heat diffusion loss rate, and corrects the flow velocity of the heat exchange tube and the filling density of the phase change material in real time.

[0010] The geological parameter identification module, including the transient thermal response test unit and the pore water pressure monitoring array, is used to identify the geological type of sand, clay or bedrock through a support vector machine classifier when the borehole radar detects abnormal fluctuations in the dielectric constant of the formation, map the heat exchange parameters, and invert the equivalent thermal conductivity and the thermal storage density correction coefficient through the thermo-permeability coupling model.

[0011] The adaptive optimization control module is used to adopt a convection-dominated heat transfer strategy when the groundwater flow velocity output by the geological parameter identification module exceeds a preset threshold or the thermal conductivity is lower than a preset threshold. The optimal control strategy is generated through a depth-deterministic strategy gradient algorithm, and the heat pump power and circulation pump frequency are adjusted in real time. At the same time, the trigger threshold of the phase change material is adjusted according to the real-time formation temperature distribution.

[0012] The global thermal balance module is used to actively compensate for heat loss by adjusting the heat pump working fluid flow rate through a nonlinear PID controller when the temperature gradient of the thermal accumulation region exceeds a preset threshold detected by the formation temperature field reconstruction module. At the same time, it dynamically adjusts the heat charge and release rate of the phase change material according to the real-time heat storage and release power ratio.

[0013] As a further technical solution of the present invention, the LA hybrid time series prediction model includes a hierarchical time series decomposition layer, a multimodal fusion layer, a dynamic weight reconstruction layer, a three-dimensional heat conduction correction layer, an anomaly frequency judgment layer, and a cross-seasonal coupling output layer;

[0014] The hierarchical temporal decomposition layer is used to generate an eigenmode function set by superimposing adaptive white noise iteratively through an improved adaptive noise set empirical mode decomposition algorithm. The layer then filters out components whose sample entropy is lower than a preset threshold of 0.8 and whose energy percentage exceeds 75%, and outputs high-frequency noise, mid-frequency diurnal periodicity and low-frequency seasonal components.

[0015] The multimodal fusion layer is used to calibrate the time series offset of meteorological and load data through a dynamic time warping algorithm. When the peak offset of the cross-correlation function between meteorological and load data exceeds a preset threshold, the time axis of meteorological data is stretched. Temperature, humidity, radiation parameters and load components are fused into a three-dimensional spatiotemporal tensor through a multi-head attention mechanism.

[0016] The dynamic weight reconstruction layer is used to mark outliers based on the KL divergence distribution of samples within a sliding window using Bayesian optimization when the Mahalanobis distance exceeds a preset confidence interval threshold. For non-outlier samples, the maximum a posteriori probability weight matrix is ​​solved using Gaussian regression. The three-dimensional heat conduction correction layer is used to construct an unstructured formation heat conduction grid model using the finite element method. When the meteorological temperature change rate ΔT / Δt > 2℃ / h, the correction coefficient matrix is ​​iteratively solved using a transient heat conduction control formula. The transient heat conduction control formula is:

[0017]

[0018] In formula (1), ρ is the soil density, with units of kg / m³. 3 C p ρ represents the specific heat capacity of the soil, in J / (kg·K), T represents the ground temperature field, t represents the time variable, in seconds, k represents the thermal conductivity of the soil, in W / (m·K), and Q represents the soil thermal conductivity.solar This is the solar radiation heat source term, with units of W / m². 3 , This is the heat flux density correction matrix;

[0019] The anomaly frequency decision layer is used to define the sequence of device start-up and shutdown events as observations based on the hidden Markov model. The implicit states include normal, transition and abnormal. If the proportion of the abnormal state Viterbi path is greater than the preset threshold, the sample pool is cleared and 30 days of historical data are reloaded.

[0020] The cross-seasonal coupling output layer is used to input the predicted heat flux density into the pile strain formula:

[0021] ε=aΔT+β+δ thermal / Δt (2)

[0022] In formula (2), ε is the strain value of the pile body, a is the thermal expansion coefficient of the pile material, ΔT is the temperature change, and β is the thermal stress coupling coefficient, with units of MPa. -1 δ thermal The value is the thermal stress, in MPa. If ε > 0.2%, the Pareto optimal solution is obtained by sequential quadratic programming algorithm, and the thermal storage demand value is output and sent to the heat pump control unit.

[0023] As a further technical solution of the present invention, the working method of the finite element inversion algorithm is as follows: Based on the coordinates of the abnormal regions marked by clustering, the target region is locally meshed using an unstructured mesh generation algorithm. Heat flux density sensor data and clustered partition temperature data are loaded, and a heat conduction forward model is used. A nonlinear least squares algorithm is employed to iteratively invert the formation thermal properties, including thermal conductivity γ and heat capacity c, to reconstruct the three-dimensional unsteady temperature field, and to calculate the heat diffusion loss rate n. loss :

[0024]

[0025] In formula (3), γ is the dynamic correction value of the formation thermal conductivity; q sensor This represents heat flux density sensor data; v is the flow rate of the working fluid inside the heat exchanger tube. Let n be the temperature gradient; if n loss If the flow rate of the heat exchange tube exceeds the preset safety threshold, the flow rate of the heat exchange tube is dynamically adjusted by the PID controller, and the optimal filling density under the current temperature field is matched by a linear interpolation algorithm based on the latent heat temperature curve of the phase change material.

[0026] As a further technical solution of the present invention, the working steps of the depth deterministic strategy gradient algorithm include:

[0027] Step 1: Based on the state-action value function, define the optimal time difference of the target policy and calculate the policy value. The formula is as follows:

[0028]

[0029] In formula (4), For the optimal time difference objective value, R t denoted as δ, where δ is the instantaneous benefit of the heat exchange control strategy at time t; Q' is the discount factor; Q' is the Q-value estimate of the target network; a' is the optimal strategy for the next step; and ω1 is the strategy smoothing factor. The action performed in the previous moment; θ Q' For the target Q network parameters; S t+1 This indicates that action a is performed at the current time t. t Then, the changes in the environmental state at time t+1; Represents the policy network parameters θ u The gradient calculation value;

[0030] Step 2: Optimize the policy network using the policy gradient update function, and calculate the gradient. The calculation formula is as follows:

[0031]

[0032] In formula (5), J(θ) u ) is the policy objective function; u(S|θ) u ) represents the policy network; ω2 is the gradient smoothing factor; Let S be the rate of change for policy updates; E represents the expected value of sampling all future states S and actions a.

[0033] Step 3: Store historical state data through an experience replay mechanism, and optimize by introducing an entropy regularization term using an adaptive entropy regularization strategy to improve the exploration capability of the heat exchange strategy. The formula is:

[0034]

[0035] In formula (6), T entropy ω is the entropy regularization loss; ω3 is the regularization coefficient; π(a) i |S) represents the policy probability distribution; N is the number of samples in the policy space;

[0036] Step 4: Calculate the optimal heat transfer control parameters by combining the adaptive adjustment of the circulating pump frequency with the nonlinear mapping of the heat pump power. As the final execution strategy.

[0037] Based on the above technical solutions, the positive and beneficial effects of the present invention are as follows:

[0038] 1. This invention dynamically adjusts the weights of load forecast samples through the LA hybrid time-series forecast model and optimizes the heat exchange control strategy in real time based on the deep deterministic strategy gradient algorithm, making the heat storage and release of energy piles more accurate. This ensures that the energy storage system adapts to the dynamic changes in building usage patterns and the uncertainty of meteorological conditions, thereby effectively reducing heat load forecasting errors, improving the supply and demand matching degree of cross-seasonal energy storage, and avoiding problems such as excessive heat storage in summer or insufficient heating in winter due to forecast deviations.

[0039] 2. The formation temperature field reconstruction module and the geological parameter identification module work together to monitor the formation temperature field and groundwater flow characteristics in real time through a distributed fiber optic temperature measurement array and a transient thermal response testing unit. Combined with a finite element inversion algorithm, the underground heat diffusion loss rate is calculated, and a support vector machine classifier identifies the geological type. This leads to the establishment of a thermodynamic-seepage coupling model to optimize heat transfer parameters, ensuring that the energy piles maintain optimal heat transfer capacity under different geological environments. This collaborative mechanism not only enhances the system's adaptability to changes in the geological environment but also reduces energy storage attenuation caused by nonlinear effects of underground heat diffusion through precise correction of formation thermal conductivity and storage density, thereby improving the long-term energy storage stability.

[0040] 3. The adaptive optimization control module and the global thermal balance module jointly construct a dynamic heat exchange regulation and thermal field equilibrium optimization mechanism. When the groundwater flow rate exceeds the standard or the formation thermal conductivity is lower than the preset threshold, the system adopts a convection-dominated heat exchange strategy to adjust the heat pump power and circulation pump frequency, and optimizes the formation heat release path according to the gradient heat release strategy, so that the heat pump working fluid flow rate and the heat charging and releasing rate of the phase change material can be dynamically adjusted according to the formation temperature. This regulation strategy effectively alleviates the formation temperature anomalies caused by insufficient heat exchange capacity or heat accumulation effect, improves the long-term operational safety of the system in complex environments, reduces the additional energy compensation requirements, and improves the overall energy efficiency. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0042] Figure 1 This is a schematic diagram of an energy pile for efficient cross-seasonal energy storage according to the present invention;

[0043] Figure 2 This is an architecture diagram of the LA hybrid time series prediction model of the present invention;

[0044] Figure 3 This is a diagram illustrating the working architecture of the finite element inversion algorithm of this invention.

[0045] Figure 4 This is a schematic diagram illustrating the working principle of the support vector machine classifier of this invention.

[0046] Figure 5 This is a diagram illustrating the working architecture of the thermo-permeation coupling model of this invention.

[0047] Figure 6 This diagram illustrates the working steps of the deep deterministic strategy gradient algorithm of this invention. Detailed Implementation

[0048] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] An efficient cross-seasonal energy storage energy pile includes: a dynamic coupling prediction module, a formation temperature field reconstruction module, a geological parameter identification module, an adaptive optimization control module, and a global thermal balance module;

[0050] In this highly efficient cross-seasonal energy storage energy pile, such as Figure 1 As shown, the cross-seasonal thermal storage demand forecast value from the dynamic coupling prediction module is output to the input of the adaptive optimization control module, serving as the decision-making basis for heat exchange regulation. Simultaneously, the building heat load forecast data from this module is transmitted to the formation temperature field reconstruction module to guide the regulation strategy of the underground temperature field. The formation temperature field reconstruction module, based on a distributed fiber optic temperature measurement array, calculates the formation temperature gradient change rate and heat diffusion loss rate, and outputs them to the global heat balance module to adjust the gradient heat release strategy. Simultaneously, the temperature distribution data from this module is input to the geological parameter identification module to assist in the correction of formation thermal conductivity. The geological parameter identification module outputs the geological type, equivalent thermal conductivity, and thermal storage density correction coefficient to the adaptive optimization control module through a support vector machine classifier to optimize heat exchange parameters. It also feeds back geological characteristic data to the formation temperature field reconstruction module to improve the accuracy of temperature field inversion. Based on the formation thermal conductivity and groundwater flow velocity data from the geological parameter identification module, the adaptive optimization control module uses a deep deterministic gradient algorithm to adjust the heat pump power and circulation pump frequency, and outputs the optimized heat exchange flow rate regulation parameters to the global heat balance module. The global thermal balance module receives the formation temperature gradient change rate, heat exchange flow rate control parameters, and heat storage-release power ratio. It uses a nonlinear PID controller to adjust the heat pump working fluid flow rate and feeds back the adjusted temperature status data to the adaptive optimization control module to achieve dynamic optimization of the system.

[0051] The dynamic coupling prediction module is used to perform multi-scale prediction of building heat load using a hybrid time-series prediction model (LA). When the building energy consumption monitoring system detects that the frequency of indoor temperature control equipment operation exceeds a threshold, the LA hybrid time-series prediction model reconstructs the weights of historical load samples through a sliding time window mechanism, while simultaneously inputting meteorological forecast data into a three-dimensional heat conduction correction coefficient matrix, and outputting a predicted value for cross-seasonal heat storage demand. Figure 2 As shown, the LA hybrid time series prediction model further includes a hierarchical time series decomposition layer, a multimodal fusion layer, a dynamic weight reconstruction layer, a three-dimensional heat conduction correction layer, an anomaly frequency decision layer, and a cross-seasonal coupling output layer.

[0052] The hierarchical temporal decomposition layer is used to generate an eigenmode function set by superimposing adaptive white noise iteratively through an improved adaptive noise set empirical mode decomposition algorithm. The layer then filters out components whose sample entropy is lower than a preset threshold of 0.8 and whose energy percentage exceeds 75%, and outputs high-frequency noise, mid-frequency diurnal periodicity and low-frequency seasonal components.

[0053] The multimodal fusion layer is used to calibrate the time series offset of meteorological and load data through a dynamic time warping algorithm. When the peak offset of the cross-correlation function between meteorological and load data exceeds a preset threshold, the time axis of meteorological data is stretched. Temperature, humidity, radiation parameters and load components are fused into a three-dimensional spatiotemporal tensor through a multi-head attention mechanism.

[0054] The dynamic weight reconstruction layer is used to mark outliers based on the KL divergence distribution of samples within a sliding window using Bayesian optimization when the Mahalanobis distance exceeds a preset confidence interval threshold. For non-outlier samples, the maximum a posteriori probability weight matrix is ​​solved using Gaussian regression. The three-dimensional heat conduction correction layer is used to construct an unstructured formation heat conduction grid model using the finite element method. When the meteorological temperature change rate ΔT / Δt > 2℃ / h, the correction coefficient matrix is ​​iteratively solved using a transient heat conduction control formula. The transient heat conduction control formula is:

[0055]

[0056] In formula (1), ρ is the soil density, with units of kg / m³. 3 C p ρ represents the specific heat capacity of the soil, in J / (kg·K), T represents the ground temperature field, t represents the time variable, in seconds, k represents the thermal conductivity of the soil, in W / (m·K), and Q represents the soil thermal conductivity. solar This is the solar radiation heat source term, with units of W / m². 3 , This is the heat flux density correction matrix;

[0057] The anomaly frequency decision layer is used to define the sequence of device start-up and shutdown events as observations based on the hidden Markov model. The implicit states include normal, transition and abnormal. If the proportion of the abnormal state Viterbi path is greater than the preset threshold, the sample pool is cleared and 30 days of historical data are reloaded.

[0058] The cross-seasonal coupling output layer is used to input the predicted heat flux density into the pile strain formula:

[0059] ε=aΔT+β+δ thermal / Δt (2)

[0060] In formula (2), ε is the strain value of the pile body, a is the thermal expansion coefficient of the pile material, ΔT is the temperature change, and β is the thermal stress coupling coefficient, with units of MPa. -1 δ thermal The value is the thermal stress, in MPa. If ε > 0.2%, the Pareto optimal solution is obtained by sequential quadratic programming algorithm, and the thermal storage demand value is output and sent to the heat pump control unit.

[0061] Among them, Adaptive Noise Ensemble Empirical Mode Decomposition (EEMD) decomposes the original building load time-series signal into multiple intrinsic mode functions (IMFs) at different time scales by superimposing adaptive white noise. During each decomposition, EEMD employs an ensemble averaging mechanism to eliminate mode aliasing. Finally, based on sample entropy calculations, dominant energy components are selected, retaining only those with sample entropy below 0.8 and an energy share exceeding 75%, to suppress short-term random fluctuations and improve long-term trend prediction accuracy. Dynamic Time Warping (DTW) is used to compare the time alignment errors between meteorological and load data. When the peak offset of the cross-correlation function exceeds a threshold, time axis stretching or compression transformation is used to align the time scales of different data sources. This method uses a dynamic programming (DP) algorithm to calculate the optimal alignment path and adjusts the meteorological data based on the cumulative distance matrix to synchronize it with the building load data, improving load prediction accuracy. The Multi-Head Attention (MHA) mechanism processes the spatiotemporal correlation between temperature, humidity, solar radiation, and building load in parallel using multiple independent attention heads. Each attention head calculates an attention score using a query-key-value (QKV) mapping, and the influence coefficients of different meteorological variables on load forecasting are obtained through weighted summation. Finally, MHA stabilizes the data distribution through residual connections and layer normalization, improving the adaptability of the load forecasting model to environmental variables. KL divergence anomaly detection technology is used to calculate the degree of change in data distribution within a sliding window to detect outliers in building load data. First, the probability density function (PDF) of the data distribution is calculated within the current window, and then the KL divergence value of this distribution relative to the previous window is calculated. If this value exceeds a preset confidence interval threshold, the data is identified as an anomaly, and anomalous data points are further filtered using Mahalanobis distance. Bayesian optimization is used to predict the optimal value of the objective function based on a probabilistic model (usually a Gaussian process) during the anomaly data correction process. During optimization, an acquisition function (AF) is used to search for the optimal solution within the data sample space. An expectation boosting (EI) strategy is used to dynamically adjust the threshold for KL divergence detection, ensuring the stability of the anomaly detection algorithm. Gaussian process regression (GPR) uses a kernel function to map the nonlinear relationships in the data and estimates the weight matrix for building load forecasting using maximum a posteriori probability (MAP). During weight reconstruction, GPR calculates the gradient of the likelihood function and uses Laplace approximation to solve for the optimal weight distribution, improving the dynamic adaptability of load forecasting. Finite element analysis (FEA) is used to simulate the thermal diffusion process of the underground temperature field. Unstructured meshes are used to divide the geological regions, and the thermal diffusion loss rate is calculated based on the transient heat conduction control equations.FEA (Functional Energy Analysis) calculates the temperature change of each grid cell through time-step iteration and corrects the formation heat conduction parameters using finite element interpolation functions, improving the accuracy of temperature prediction. Hidden Markov Models (HMMs) construct a hidden state-observation probability transition matrix to analyze the state changes of building temperature control equipment during start-up and shutdown events. This model calculates the optimal hidden state path based on the Viterbi algorithm and determines whether to clear the sample pool and reload historical data based on the proportion of abnormal states, thereby improving the stability of load prediction. Sequential Quadratic Programming (SQP) uses a second-order optimization method to solve for the Pareto optimal solution of cross-seasonal thermal storage demand. This method calculates the optimization direction through Lagrange dual optimization and solves a quadratic approximate subproblem in each iteration to ensure optimal control of thermal storage demand when pile thermal strain exceeds limits, improving the long-term stability of the building energy storage system.

[0062] In engineering applications, the implementation of this solution follows these steps: First, a building energy consumption monitoring system is used to collect real-time start / stop signals from indoor temperature control equipment, meteorological parameters (temperature, humidity, radiation), and pile strain data. A raw database is constructed with a 1-minute granularity, and a sliding time window (7 days window, 1-day step) is used to divide the training and validation sets. Second, the CEEMDAN algorithm is used to perform modal decomposition on historical load data, extracting low-frequency seasonal components as the baseline for cross-seasonal heat storage demand. A dynamic time warping algorithm is used to align the meteorological and load time series phases. A multi-head attention network is implemented using the PyTorch framework to generate a multimodal tensor that integrates three-dimensional spatiotemporal features. Subsequently, KL divergence and Mahalanobis distance are calculated using Scikit-learn, and Bayesian optimization is performed using the GPyOpt toolkit to remove outliers. Gaussian process regression is then used to dynamically allocate the weight matrix. Simultaneously, a finite element model of ground heat conduction is constructed in COMSOL Multiphysics, embedding a Python script to update the heat flux density correction coefficient in real time, ensuring model accuracy under sudden meteorological changes. The anomaly handling process employs the hmmlearn library to train a Hidden Markov Model, setting an anomaly threshold (30%). When an anomaly is triggered, an SQL script is invoked to clear the sample pool in the Redis cache and reload historical data from the InfluxDB time-series database. Finally, the optimized thermal storage demand value is sent to the heat pump PLC controller via the Modbus protocol. Real-time monitoring of pile strain sensor data is performed; if thermal stress exceeds the limit (ε > 0.2%), the IPOPT solver is invoked to recalculate the Pareto optimal solution, dynamically adjusting the working fluid flow rate and the heat release / charge rate of the phase change material to achieve efficient and safe operation of the energy storage system.

[0063] Compared to existing technologies, this solution significantly improves the accuracy of cross-seasonal thermal storage demand forecasting through hierarchical temporal decomposition and multimodal fusion, enhances the model's generalization ability by combining dynamic weight optimization and thermal conduction physical correction; the anomaly judgment mechanism effectively avoids system-level disturbance risks, and the thermo-mechanical coupling control strategy balances structural safety and energy storage efficiency, thus improving the long-term operational stability and energy utilization rate of energy piles and providing reliable technical support for low-carbon and intelligent energy supply in buildings.

[0064] In implementation, the hardware environment for the dynamic coupling prediction module includes:

[0065] Temperature control equipment monitoring sensors (installed in the building's air conditioning / underfloor heating circuits to collect start-stop frequencies); weather station (located on the building roof, including temperature, humidity, and radiation sensors); pile heat flow meter (embedded in the energy pile at depths of 5m, 10m, and 15m underground to monitor heat flux density); strain sensor (attached to the pile surface to detect axial / radial strain); data acquisition module (centrally processes sensor signals and transmits them to the server via RS485); heat pump PLC controller (connects the working fluid circulation pump and the phase change material heat storage unit).

[0066] Based on the above hardware environment, a comparative experiment was designed. Group A used a dynamic coupling prediction module, while Group B used an ARIMA + static model; that is, the ARIMA time series model was combined with a static thermal conductivity coefficient (fixed value 0.8W / (m·K)), ignoring dynamic weight optimization and anomaly detection mechanisms. Groups A and B shared the same sensors and heat pump equipment, recording the heat load prediction error, equipment start-up and shutdown times, pile strain values, energy consumption, and thermal storage efficiency every 5 minutes; the sliding window for Group A was 7 days, and for Group B, the ARIMA order (p,d,q) was (3,1,2), with a fixed thermal conductivity coefficient; the outdoor temperature fluctuated between -5℃ and 5℃ during the experiment, and the building heat load demand was constant (30kW); the experimental data are shown in Table 1.

[0067] Table 1. Module Comparison Experiment Record Table

[0068]

[0069] Experimental results show that the heat load prediction error of group A (RMSE = 1.24kW) is significantly lower than that of group B (3.86kW), due to the improved generalization ability of the model achieved by dynamic weight optimization and multimodal fusion; the start-stop frequency of group A (2.10 times / hour) is 57.3% lower than that of group B (4.92 times / hour), indicating that the dynamic anomaly judgment mechanism effectively suppresses ineffective regulation; the maximum strain value of the pile body in group A (0.164%) is lower than that in group B (0.290%), indicating that three-dimensional heat conduction correction and Pareto optimization alleviate thermal stress accumulation; the average daily energy consumption of group A (58.92kWh) is 21.4% lower than that of group B (74.98kWh), and the thermal storage efficiency is improved by 26.3% (82.08% vs 64.92%). Therefore, the dynamic coupling prediction module significantly improves the prediction accuracy, operational stability, and energy efficiency of energy piles through multi-scale optimization and physical correction.

[0070] The formation temperature field reconstruction module is used to collect formation temperature field gradient data and axial and radial temperature gradients within the energy pile using a distributed fiber optic temperature measurement array. When the rate of change of the formation temperature gradient exceeds a preset critical value, it reconstructs the three-dimensional unsteady temperature field based on the heat flux density sensor and temperature sensor data within the energy pile, using a finite element inversion algorithm. It calculates the heat diffusion loss rate and corrects the flow velocity of the heat exchanger tubes and the phase change material filling density of the underground thermal storage body in real time. Further, such as... Figure 3 As shown, the working method of the finite element inversion algorithm is as follows: Based on the coordinates of the anomalous regions marked by clustering, the target region is locally meshed using an unstructured mesh generation algorithm. Heat flux density sensor data and clustered partition temperature data are loaded, and a heat conduction forward model is used. A nonlinear least squares algorithm is employed to iteratively invert the formation's thermal properties, including thermal conductivity γ and heat capacity c, to reconstruct the three-dimensional unsteady temperature field and calculate the heat diffusion loss rate n. loss :

[0071]

[0072] In formula (3), γ is the dynamic correction value of the formation thermal conductivity; q sensor This represents heat flux density sensor data; v is the flow rate of the working fluid inside the heat exchanger tube. Let n be the temperature gradient; if n loss If the flow rate of the heat exchange tube exceeds the preset safety threshold, the flow rate of the heat exchange tube is dynamically adjusted by the PID controller, and the optimal filling density under the current temperature field is matched by a linear interpolation algorithm based on the latent heat temperature curve of the phase change material.

[0073] The unstructured mesh generation algorithm is used for high-precision discretization of the formation temperature field. First, anomalous temperature gradient regions are identified based on clustering markers. A density peak clustering method is employed to determine the center points of anomalous regions based on the local density and relative distance of the temperature data, generating cluster partitions. Subsequently, Delaunay triangulation is used to generate locally high-density meshes within anomalous regions, while maintaining a uniform mesh distribution in normal regions to reduce computational load and improve finite element solution accuracy.

[0074] The forward model of thermal conduction is used to simulate the thermal diffusion behavior of the formation and serves as the forward computational framework for finite element inversion. Based on the Fourier heat conduction equation, it describes the relationship between heat flux density, thermal conductivity, and temperature gradient, and constructs time-dependent thermal diffusion control equations. During the calculation, the implicit finite difference method (IFDM) is used for time step control, and adaptive boundary conditions are combined to correct the heat exchange intensity at the heat transfer pile-soil interface, ensuring that the simulated temperature field accurately reflects the actual thermal diffusion of the formation.

[0075] A nonlinear least squares inversion algorithm is used to solve for the thermal properties of the formation, including thermal conductivity and specific heat capacity. First, based on the Levenberg-Marquardt optimization (LM optimization) method, an objective function is constructed to minimize the mean square error (MSE) between the simulated temperature and the sensor-measured temperature. Second, a multi-scale iterative regularized MSIR method is combined to prevent parameter overfitting or convergence to a local optimum during the inversion process. Finally, the partial derivatives are solved using Newton's iteration method, and the parameter update step size is optimized to achieve an optimal balance between convergence accuracy and computational efficiency in the inversion calculation.

[0076] The thermal diffusion loss rate is used to quantify the heat transfer performance of underground thermal reservoirs. First, the heat dissipation mode during the heat transfer process is analyzed based on the Laplace transform, and the heat flow loss of different grid cells is calculated using the finite volume method (FVM). Subsequently, entropy increase analysis is used to assess the irreversible energy loss during formation thermal diffusion, and adaptive temperature gradient threshold control is used to ensure that the thermal diffusion loss rate is maintained within a safe range.

[0077] In the implementation of this module: fiber optic temperature measurement nodes are arranged every 0.5m along the axial direction inside the pile foundation, and a ring-shaped temperature measurement array is set radially around the pile body; heat flux density sensors (range 0~1000W / m) are installed on the outer wall of the heat exchange pipeline. 2), surface weather stations are deployed to monitor temperature, humidity, and solar radiation data. Temperature gradients, heat flux density, and meteorological parameters are collected in real time. Temperature change rate is extracted using a sliding time window (24h window length), and K-means clustering algorithm is used to identify anomalous areas (standard deviation σ ≥ 1.2℃). The Delaunay mesh generator is used to locally refine the anomalous areas, a forward model is loaded, and a finite element inversion algorithm is used to iteratively correct γ and c until the temperature residual converges. When n loss When the flow rate is ≥10%, the PID controller outputs a heat exchanger tube flow rate adjustment command (v=0.51200kg / m). 3 The data is sent in real time to the circulating pump and phase change material injection system. If the inversion iterations fail to converge after more than 50 iterations, the mesh reconstruction process is automatically triggered to reset the local encrypted region and reload the data.

[0078] Compared to traditional static temperature field monitoring methods, this module achieves high-resolution thermal field reconstruction through distributed optical fibers and dynamic inversion algorithms. Combined with PID and phase change material optimization strategies, it significantly reduces heat diffusion losses and improves thermal storage efficiency. Local grid refinement in anomalous areas and inversion of unsteady-state parameters enhance the model's adaptability to changes in formation thermal properties, effectively avoiding the risks of overheating or heat leakage, and providing accurate and reliable thermal management capabilities for cross-seasonal energy storage systems.

[0079] To verify the effectiveness of the finite element inversion algorithm in cross-seasonal energy storage piles, a comparative experiment was designed: Group A used the finite element inversion algorithm, while Group B used a traditional static empirical model (with a fixed thermal conductivity λ = 1.5 W / m·K, and the temperature field estimated based on linear interpolation). The experiments were conducted on the same sand-clay alternating strata site. Both groups deployed distributed fiber optic temperature measurement arrays, heat flux density sensors, and circulating pump systems, with a constant heat flux density of 500 W / m². 2 The system continuously collected temperature field data, heat diffusion loss rate, and energy consumption every 30 minutes for 24 hours. Group A used Delaunay mesh local refinement and Levenberg-Marquardt optimization to retrieve thermal conductivity and heat capacity, dynamically adjusting the heat exchanger tube flow velocity (0.5-2.0 m / s) and phase change material packing density (800-1200 kg / m³). 3 Group B operates according to a fixed strategy (flow rate 1.0 m / s, filling density 1000 kg / m³). 3 The experimental record table is shown in Table 2:

[0080] Table 2. Experimental Record of BP Finite Element Inversion Algorithm

[0081]

[0082] Experimental results show that Group A, which uses the finite element inversion algorithm, is significantly better than Group B, which uses the traditional static model, in terms of core indicators such as temperature field reconstruction accuracy, thermal conductivity inversion error, and thermal diffusion loss rate.

[0083] The geological parameter identification module includes a transient thermal response testing unit and a pore water pressure monitoring array. When borehole radar detects abnormal fluctuations in the formation's dielectric constant, it uses a support vector machine classifier to identify the geological type (sand, clay, or bedrock), maps heat exchange parameters, and inverts the equivalent thermal conductivity and storage density correction coefficient using a thermodynamic-permeability coupling model. The transient thermal response testing unit applies a short-duration thermal pulse to the formation surrounding the borehole using pulsed heat flow excitation and collects formation temperature response data using a fiber Bragg grating sensor to calculate the transient thermal diffusion coefficient and identify the initial thermal properties of the formation. The pore water pressure monitoring array extracts groundwater flow trends using a dynamic water pressure gradient analysis method and calculates soil permeability parameters using adaptive permeability regression (APR) to obtain correction factors for the influence of groundwater on heat diffusion. Further, such as... Figure 4 As shown, the working method of the support vector machine classifier is as follows: First, the dielectric properties data of the formation are projected to a high-dimensional feature space through kernel function mapping. Then, the Lagrange dual optimization method is used to maximize the data class margin, and soft-margin support vectors are used to enhance the adaptability of the model to geological transition zones. During the classification process, the support vector machine classifier uses KKT conditional filtering to dynamically update the support vectors and remove redundant data points. Subsequently, the support vector weight distribution is optimized through an adaptive weight adjustment mechanism so that the decision boundary of the classifier between sand, clay and bedrock categories conforms to the trend of formation thermal property changes. Finally, the support vector machine classifier uses a heat exchange parameter mapping mechanism to call the corresponding thermal diffusivity, heat capacity ratio and permeability from a predefined thermal property parameter library according to the classification results.

[0084] The transient thermal response testing unit employs pulsed heat flow excitation technology to apply short-duration thermal pulses to the formation surrounding the borehole, thereby stimulating the formation's transient thermal response behavior. Fiber Bragg grating sensors record the formation temperature variation curve over time, and a time-temperature response model is used to calculate the formation's transient thermal diffusivity. During data processing, a heat flux density correction algorithm is used to compensate for surface heat loss, and a partial differential temperature gradient inversion method is combined to optimize measurement accuracy, accurately identifying the formation's initial thermal properties, including thermal conductivity, specific heat capacity, and thermal diffusivity. The pore water pressure monitoring array obtains groundwater flow trends using a dynamic water pressure gradient analytical method. First, the transient water level disturbance response is used to analyze the groundwater permeability rate, and local permeability changes are calculated based on the Poisson equation. Subsequently, an adaptive permeability regression model is used to dynamically fit the permeability distribution within multiple time windows to obtain a correction factor for the impact of groundwater flow on formation thermal diffusivity. Finally, a multi-scale permeability correction mechanism is used to adjust permeability parameters, enabling the calculation results to adapt to dynamic changes under different geological conditions.

[0085] A Support Vector Machine (SVM) classifier is used to identify formation types and map corresponding heat transfer parameters based on formation dielectric property data obtained from borehole radar. First, a kernel function mapping method is employed to project the low-dimensional dielectric property data into a high-dimensional feature space to enhance the linear separability of the data. Then, Lagrange dual optimization is used to solve the objective function of the SVM, maximizing the hyperplane margin of data classes. This is combined with a soft-margin SVM enhancement strategy to improve the classifier's adaptability to geological transition zones. During classification, support vectors are dynamically updated based on KKT conditions, and redundant data points are removed to reduce computational complexity. Next, adaptive weight adjustment optimizes the support vector weights so that the classification boundary conforms to the changing trends of formation thermal properties. Finally, the classification results are automatically matched with corresponding thermal diffusivity, heat capacity ratio, and permeability by calling a predefined thermal property parameter library through a heat exchange parameter mapping mechanism to optimize the heat transfer performance of the energy storage system.

[0086] In the specific application of high-efficiency cross-seasonal energy storage piles, those skilled in the art can achieve geological parameter identification function by following these steps:

[0087] Fiber Bragg grating temperature sensors (FBG-T) and pressure sensors (FBG-P) were deployed axially at 1m intervals within the borehole. A transient thermal response testing unit (2kW power, adjustable pulse width) and a data acquisition terminal were installed on the surface. The transient thermal response testing unit applied thermal pulses to the formation, and the FBG-T recorded the temperature response curve at a sampling rate of 1kHz. The temperature decay characteristics were extracted using Fast Fourier Transform (FFT). A pore water pressure monitoring array acquired water pressure gradient data in real time, and moving average filtering was used to eliminate noise, generating a dynamic water pressure gradient time series. Next, the temperature decay curve was fitted based on a non-Fourier heat conduction model to invert the initial thermal conductivity and heat capacity ratio. The permeability and non-Darcy flow coefficient were solved iteratively using the Darcy-Forchheimer equation to output the groundwater correction factor. An RBF kernel SVM model was pre-trained using dielectric constant and loss tangent as feature vectors to optimize the soft-spacing parameters and kernel width. When the borehole radar detects abnormal fluctuations in the dielectric constant, it triggers the SVM classifier, inputting a real-time feature vector and outputting the probability of sand, clay, or bedrock categories. Based on the classification results, it calls predefined thermophysical parameters, inputting the equivalent thermal conductivity and storage density correction coefficient from the thermodynamic-permeability coupling model. Finally, the equivalent thermal conductivity and storage density correction coefficient from the thermodynamic-permeability coupling model are sent to the storage body control unit to adjust the flow rate of the heat exchanger tubes and synchronously update the groundwater correction factor to the heat diffusion loss rate calculation module.

[0088] like Figure 5 As shown, the workflow of the thermo-seepage coupling model begins with the input of the classified geological type. First, it constructs the unsteady thermo-seepage coupling equations (combining the energy equation and Darcy's law). If the finite element multi-scale inversion method (FEMI) successfully discretizes the temperature field-velocity-heat transfer matrix, it enters the Poisson equation constraint (PEC) correction stage. If discretization fails, the energy equation is re-parameterized and reconstructed. In the velocity correction stage, when the error in the Poisson-corrected water velocity exceeds a 2% threshold, the coupling coefficient of Darcy's law is automatically adjusted and cyclically corrected until the accuracy requirements are met. The correction parameters are verified through sensitivity analysis. If the error exceeds 5%, the geological type reclassification process is triggered, and the model inversion is restarted. If the verification passes, the parameters are pushed to the reservoir control system to dynamically adjust the operating strategy. The entire process forms a closed loop through discretization fault tolerance, velocity correction loops, and parameter feedback mechanisms, ensuring dynamic matching between the model output and the formation thermal properties.

[0089] In practice, this invention significantly improves the accuracy and robustness of formation thermal property parameter identification by fusing multi-source data from transient thermal response testing and pore water pressure monitoring, combined with dynamic optimization decision-making using a support vector machine classifier. Compared to traditional empirical models, it can accurately identify sand-clay transition zones and bedrock strata, dynamically correct thermal conductivity and permeability parameters, effectively suppress the interference of groundwater seepage on heat diffusion, and optimize thermal storage density control strategies. This enhances the geological adaptability and long-term operational stability of cross-seasonal energy storage systems, providing technical support for efficient thermal energy management under complex geological conditions.

[0090] To verify the advantages of the geological parameter identification module (Group A) over the traditional geological experience model (Group B) in terms of stratigraphic classification accuracy and thermal property parameter inversion, a comparative experiment was conducted. Group B adopted an empirical classification method based on borehole core samples (manually identifying sand / clay / bedrock), combined with fixed thermal conductivity (sand λ = 1.5 W / m·K, clay λ = 1.2 W / m·K, bedrock λ = 2.8 W / m·K) and permeability (k = 1e-4 m / s) parameters, without a dynamic correction mechanism.

[0091] The experiment involved two sets of pile foundations (Group A and Group B) in the same area, with alternating layers of sand, clay, and bedrock (thickness 0.5–2 m); the heat source input was a constant heat flux density of 400 W / m³. 2 The experiment lasted for 48 hours; the experimental data are recorded in Table 3.

[0092] Table 3 Experimental Record of Geological Parameter Identification Module

[0093]

[0094] As shown in Table 3, the geological parameter identification module (Group A) significantly outperforms the traditional empirical model (Group B) in key indicators such as stratigraphic classification accuracy (improved by 32.9%), thermal conductivity inversion error (reduced by 79.0%), and permeability correction error (reduced by 75.6%). Through dynamic classification using support vector machines and coupled inversion with thermal seepage, it accurately identifies sand-clay transition zones and bedrock strata, adaptively corrects thermal property parameters, reduces heat diffusion loss rate (reduced by 58.4%) and system energy consumption (reduced by 30.4%), while simultaneously improving temperature field stability (σ reduced by 64.3%) and phase change material filling accuracy (error reduced by 79.3%). Experiments demonstrate that this module can effectively address the challenges of thermal property parameter identification under complex stratigraphic conditions, providing reliable technical support for the efficient operation of cross-seasonal energy storage systems.

[0095] An adaptive optimization control module is used to employ a convection-dominated heat transfer strategy when the groundwater flow velocity output by the geological parameter identification module exceeds a preset threshold or the thermal conductivity falls below a preset threshold. This strategy generates the optimal control strategy through a deep deterministic strategy gradient algorithm, adjusting the heat pump power and circulation pump frequency in real time. Simultaneously, it adjusts the trigger threshold of the phase change material based on the real-time formation temperature distribution. Specifically, the module works as follows: based on the groundwater flow velocity and formation thermal conductivity anomaly signals from the geological parameter identification module, constraint conditions are established through a boundary constraint mechanism. Subsequently, the heat pump power and circulation pump frequency are optimized and adjusted using the convection-dominated heat transfer strategy. During the control process, a multi-step time-series strategy is used to evaluate and calculate the cumulative return value under different heat transfer strategies, and the control parameters are corrected using a strategy gradient adaptive update method. Next, the optimal trigger temperature of the phase change material is calculated based on a real-time formation temperature distribution prediction method, and the phase change thermal storage process is dynamically adjusted using a phase change trigger threshold dynamic adjustment mechanism. Finally, deep reinforcement learning feedback optimization iteratively adjusts the heat transfer parameters to optimize the long-term thermal storage stability of the energy pile. Further, as... Figure 6 As shown, the working steps of the depth deterministic policy gradient algorithm include:

[0096] Step 1: Based on the state-action value function, define the optimal time difference of the target policy and calculate the policy value. The formula is as follows:

[0097]

[0098] In formula (4), For the optimal time difference objective value, R t denoted as δ, where δ is the instantaneous benefit of the heat exchange control strategy at time t; Q' is the discount factor; Q' is the Q-value estimate of the target network; a' is the optimal strategy for the next step; and ω1 is the strategy smoothing factor. The action performed in the previous moment; θ Q' For the target Q network parameters; S t+1 This indicates that action a is performed at the current time t. t Then, the changes in the environmental state at time t+1; Represents the policy network parameters θ u The gradient calculation value;

[0099] Step 2: Optimize the policy network using the policy gradient update function, and calculate the gradient. The calculation formula is as follows:

[0100]

[0101] In formula (5), J(θ) u ) is the policy objective function; u(S|θ) u ) represents the policy network; ω2 is the gradient smoothing factor; Let S be the rate of change for policy updates; E represents the expected value of sampling all future states S and actions a.

[0102] Step 3: Store historical state data through an experience replay mechanism, and optimize by introducing an entropy regularization term using an adaptive entropy regularization strategy to improve the exploration capability of the heat exchange strategy. The formula is:

[0103]

[0104] In formula (6), τ entropy ω is the entropy regularization loss; ω3 is the regularization coefficient; π(a) i |S) represents the policy probability distribution; N is the number of samples in the policy space;

[0105] Step 4: Calculate the optimal heat transfer control parameters by combining the adaptive adjustment of the circulating pump frequency with the nonlinear mapping of the heat pump power. As the final execution strategy.

[0106] The boundary constraint control first determines whether the groundwater flow velocity and thermal conductivity data from the geological parameter identification module exceed the set heat transfer adaptation threshold. The system employs the constrained Lagrange multiplier method to establish heat transfer boundary constraints and constructs a solution method for the optimal solution of the penalty term to limit heat transfer parameters exceeding the safe range. Simultaneously, a gradient descent-based boundary search strategy ensures that heat transfer parameters maintain an optimal distribution within the feasible solution range, guaranteeing that the heat transfer process will not lead to energy storage imbalance due to fluctuations in geological characteristics. The system monitors the formation temperature field and groundwater flow velocity data in real time and uses the fluid momentum balance equation to calculate the dominant heat transfer mode in the formation. Under high groundwater flow velocities, the system prioritizes a convection heat transfer enhancement strategy and combines it with a flow field adaptive heat transfer adjustment model to calculate the optimal allocation scheme for heat pump power. Simultaneously, an adaptive control method based on the heat transfer medium temperature gradient dynamically adjusts the circulation pump frequency to optimize fluid heat transfer efficiency. In low flow velocity environments, the system switches to a conduction-dominated mode, optimizing the heat storage time based on the formation heat diffusion equation to improve the uniformity of underground heat storage.

[0107] When evaluating time-series strategies, the system employs a multi-step time-series reward calculation method based on reinforcement learning to assess the long-term benefits under different heat exchange control strategies. First, a state-action value function is constructed, and the expected reward of the current heat exchange strategy is calculated using a discounted reward optimization model. During optimization, the system uses a time-series data smoothing strategy to perform regression analysis on heat exchange data within different historical time steps to improve the stability of strategy optimization. Simultaneously, a dynamic step-size update mechanism adjusts the weights of the heat exchange strategy based on changes in groundwater flow velocity to ensure that the system's heat exchange control conforms to the long-term optimal energy distribution.

[0108] During strategy optimization, the system employs a strategy gradient update mechanism to calculate the impact of heat transfer strategies on overall returns. First, based on a Gaussian distribution strategy sampling method, different heat transfer parameters are selected within the strategy space for exploration, and their heat transfer efficiencies are calculated. Then, the optimal update direction of the strategy is calculated using a gradient approximation method, and combined with an adaptive learning rate adjustment mechanism, the strategy update magnitude is dynamically optimized at each time step. Furthermore, the system also employs a strategy pruning method based on optimal solution constraints to eliminate non-optimal heat transfer strategies, thereby improving optimization efficiency.

[0109] Adaptive threshold control is based on a real-time formation temperature distribution prediction method to calculate the optimal trigger temperature for the phase change material (PCM). When temperature changes drastically, the system employs a phase change thermodynamic equilibrium model to optimize the PCM packing density and combines this with a nonlinear adjustment strategy for the PCM energy storage threshold to dynamically correct the PCM storage process under different thermal storage requirements. Simultaneously, the system uses a packing optimization method based on the principle of minimum thermal resistance to adjust the spatial distribution structure of the PCM, thereby improving overall thermal storage efficiency.

[0110] The system continuously optimizes heat transfer parameters under different underground heat transfer environments based on a reinforcement learning feedback optimization mechanism. First, it stores heat transfer data under different operating conditions using an experience replay mechanism, and combines this with an adaptive pruning method for the reinforcement learning network structure to optimize the complexity of the learning model and improve computational efficiency. During heat transfer control, the system searches for the optimal policy within the heat transfer policy space using a cross-entropy-based policy exploration method, and combines this with a dynamic convergence mechanism based on a policy adversarial network to ensure that the heat transfer control parameters consistently converge to the optimal solution under different environments.

[0111] In practical implementation, the implementation process of this module is as follows: First, a distributed fiber optic temperature measurement array and thermal conductivity sensor are deployed in the formation to collect groundwater flow velocity and thermal conductivity data in real time. The control module is triggered when the flow velocity is greater than 0.1 m / s or the thermal conductivity is less than 1.2 W / (m·K). A DDPG network is built using the TensorFlow framework. The Actor network input includes heat pump power, circulation pump frequency, formation temperature distribution, and phase change material state. The output is the heat pump power increment (range ±10%) and circulation pump frequency adjustment (±5 Hz). The Critic network evaluates the state action values ​​and generates the Q-value gradient. During the training phase, an experience replay pool is used to store 100,000 sets of historical data with a batch size of 256. The learning rate is set to Actor 1e-4 and Critic 3e-4, and the entropy regularization coefficient is initially set to 0.2 and decays with training. In real-time control, formation temperature field data is collected every 5 minutes. The trained strategy network is then used to generate control commands, which are transmitted to the heat pump inverter and circulating pump PLC via the Modbus protocol. The phase change material trigger threshold is adjusted synchronously (error tolerance ±0.5℃). In the anomaly handling phase, if the heat pump power exceeds the limit (>120% of rated value) or the frequency overshoots (>50Hz), the system switches to PID safety control mode and triggers online fine-tuning of the strategy network. During system maintenance, the experience pool data is updated monthly, and transfer learning is used to adapt to seasonal geological parameter changes.

[0112] Compared to traditional PID and static threshold control methods, this module achieves adaptive optimization of heat exchange strategies through deep reinforcement learning, significantly improving the thermal storage stability in complex geological environments. The dynamic phase change triggering mechanism and multi-objective benefit function design effectively balance energy efficiency and equipment lifespan, avoiding thermal stress concentration and over-adjustment, and providing highly robust intelligent control capabilities for cross-seasonal energy storage systems.

[0113] The adaptive optimization control module of this invention is applied to energy piles for high-efficiency cross-seasonal energy storage. Its hardware environment includes: borehole radar monitoring changes in the dielectric properties of the formation, combined with fiber Bragg grating (FBG) temperature sensors and a pore water pressure monitoring array to collect real-time data on groundwater flow velocity, formation thermal conductivity, and temperature gradient. All sensors are arranged along the energy pile to achieve three-dimensional monitoring of the formation thermal field. For the heat exchange system, a high-efficiency dual-source heat pump unit is configured, forming a complete underground heat exchange loop with the heat exchange pipe network and circulating pump. Phase change energy storage material modules are filled inside the energy pile to optimize heat storage and regulation capabilities. The data acquisition and control section consists of an industrial-grade programmable logic controller (PLC), an edge computing unit (ECU), and a deep reinforcement learning control server. The PLC is used for real-time heat exchange control, the ECU is responsible for data parsing and algorithm execution, and the control server runs a deep deterministic policy gradient algorithm (DDPG) and performs optimization calculations based on real-time data. The entire system is remotely monitored and controlled by a SCADA system to ensure long-term stable operation of the energy pile in complex environments.

[0114] Based on the above hardware operating conditions, experiments were conducted in an underground thermal storage test field. Two sets of high-efficiency inter-seasonal energy piles of the same specifications were selected, and heat exchange performance was compared using an adaptive optimization control module (Group A) and a traditional PID control method (Group B) to verify the advantages of the present invention in practical applications. Group A used a deep deterministic strategy gradient algorithm (DDPG) to optimize the heat exchange strategy, adaptively adjusting the heat pump power, circulation pump frequency, and phase change material thermal storage threshold based on real-time geological data. Group B used fixed-gain PID control, adjusting heat exchange parameters based on a preset target temperature, but it could not adapt to dynamic changes in geological conditions. During the experiment, identical temperature, flow, and pressure sensors were placed around both sets of energy piles to ensure data comparability, and the same initial formation temperature was set before the experiment. Two typical geological environments (sand area with high groundwater flow velocity and clay area with low flow velocity) were selected to test the adaptability of the two control strategies under different heat exchange conditions. The experimental record table is shown in Table 4.

[0115] Table 4. Experimental Record of Adaptive Optimization Control Module Application

[0116]

[0117] Experimental results show that Group A energy piles using the adaptive optimization control module are significantly better than Group B using traditional PID control in terms of heat exchange response speed, heat storage efficiency, and formation temperature stability.

[0118] The global thermal balance module is used to actively compensate for heat loss by adjusting the heat pump working fluid flow rate through a nonlinear PID controller when the temperature gradient of the thermal accumulation region exceeds a preset threshold detected by the formation temperature field reconstruction module. At the same time, it dynamically adjusts the heat charge and release rate of the phase change material according to the real-time heat storage and release power ratio. The working principle of the gradient heat release strategy is as follows: Based on the temperature gradient data from the formation temperature field reconstruction module, the spatial distribution of the heat accumulation region is calculated using a multi-scale thermal zoning clustering method, and the heat diffusion path is identified using the gradient direction field fitting method GDFF to generate a dynamic heat release priority weight matrix. Next, based on the dynamic feedback of the formation thermal diffusivity and the heat pump working fluid flow rate, a nonlinear PID heat flow control method is used to adjust the heat exchange flow rate in real time. This nonlinear PID heat flow control method suppresses temperature fluctuations caused by unsteady heat transfer through adaptive disturbance compensation. Subsequently, the HSRP-AA method is used to calculate the deviation between the current heat storage rate of the phase change material and the formation heat release demand, and the PCD-CDR method is called to optimize the heat release rate of the phase change material, ensuring that the heat release matches the local temperature gradient change trend. Finally, based on the formation heat release rate, working fluid flow rate fluctuations, and heat storage stability, the heat pump flow control strategy is adaptively updated using the MOTB-IO method, a multi-objective thermal balance iterative optimization method.

[0119] The multi-scale thermal zoning clustering method is used to calculate the spatial distribution of thermal accumulation regions. First, based on the adaptive K-means clustering method, cluster analysis is performed on the temperature gradient data output by the formation temperature field reconstruction module. Local density peak detection is used to identify high-temperature zones, and a multi-scale data fusion strategy is employed to optimize the cluster boundaries, ensuring consistency among temperature anomaly regions at different scales. Subsequently, the system constructs a formation thermal accumulation feature space based on thermal diffusion timescale mapping, forming a three-dimensional distribution model of thermal accumulation regions for dynamic adjustment of subsequent heat release strategies. The gradient direction field fitting method is used to identify thermal diffusion paths to optimize the heat release sequence. First, based on the finite difference thermal gradient calculation method, a direction field of formation temperature changes is constructed, and a Gaussian curvature constraint fitting algorithm is used to calculate the gradient change trend. Next, an adaptive Laplace smoothing operator is used to optimize the gradient direction to eliminate abnormal gradient points and improve the accuracy of thermal diffusion direction prediction. Finally, through dynamic heat release priority calculation, the thermal diffusion paths of different regions are weighted to form a heat release priority weight matrix. The nonlinear PID heat flow control method is used to optimize the heat exchange flow rate based on dynamic feedback from the formation thermal diffusivity and the heat pump working fluid flow rate. First, based on a variable-gain PID control strategy, a temperature-flow dual-parameter adaptive adjustment model is adopted to dynamically adjust the PID gain parameter with temperature changes. Then, disturbance suppression compensation control is used to adaptively suppress temperature fluctuations caused by unsteady heat transfer by calculating the flow regulation overshoot in real time. The system further incorporates a second-order differential lead compensation mechanism to smooth the heat flow control command, improving the response stability of the heat exchange system. The heat storage-to-release power ratio adjustment method optimizes the heat storage and release process by calculating the deviation between the phase change material's heat storage rate and the formation's heat release demand. The system first calculates the real-time heat transfer power of the phase change material based on the heat storage-to-release ratio calculation model and matches it with the formation's heat release demand. Subsequently, a power mismatch least-squares optimization method is used, introducing a dynamic thermal resistance balance factor into the objective function to optimize the phase change material's charge and release rate, allowing it to dynamically adapt to the changing trend of the formation's temperature distribution. The phase change dynamic charge and release control method is used to optimize the phase change material's heat release rate, matching it with local temperature gradient changes. The system is based on a thermodynamic non-equilibrium phase change model to calculate the latent heat release characteristics of the phase change material (PCM) under different temperature fields, and uses a dynamic optimal charge / release heat algorithm to adjust the charge / release heat rate of the PCM. Subsequently, a linear interpolation control strategy is adopted to optimize the charge / release heat curves in different temperature ranges to improve the heat transfer efficiency of the PCM. A multi-objective thermal balance iterative optimization method is used to optimize the heat pump flow control strategy based on the formation heat release rate, working fluid flow fluctuation, and thermal storage stability. First, a multi-objective optimization function is constructed, taking heat transfer stability, energy utilization efficiency, and formation thermal diffusion uniformity as optimization objectives, and a multi-objective optimization algorithm based on Pareto front search is used to solve for the optimal combination of heat transfer parameters.Secondly, by using an adaptive weight adjustment mechanism, the weight ratios of different optimization objectives are adjusted in real time, enabling the heat exchange strategy to dynamically adapt to changes in the external environment and improve the long-term heat storage balance of the system.

[0120] In engineering applications, the implementation of this module follows these steps: First, a distributed fiber optic temperature measurement array (0.5m spacing) and heat flux density sensor are deployed along the axial and radial directions of the energy pile to collect real-time ground temperature gradient data. When a local temperature gradient > 3℃ / m is detected, a gradient heat release strategy is triggered. Multi-scale thermal partitioning clustering is performed using Python's Scikit-learn library to divide the thermally accumulated sub-regions, and the gradient direction field fitting plugin (GDFFToolbox) of COMSOL Multiphysics is called to generate a heat release priority matrix. A nonlinear PID controller is built in MATLAB / Simulink, inputting the thermal diffusivity coefficient and working fluid flow feedback signals, and outputting heat pump inverter adjustment commands (frequency range 25-50Hz). The phase change material charge / discharge heat control stage uses the HSRP-AA algorithm written in LabVIEW. The temperature-latent heat curve of the phase change material is read via the OPC UA protocol, and the charge / discharge rate deviation is calculated in real-time and sent to the phase change unit controller. The multi-objective optimization module integrates the NSGA-II algorithm (population size 100, iterations 500), performing global optimization every 6 hours to generate joint control parameters for the heat pump and phase change material. An anomaly handling mechanism is configured to switch to safe mode and trigger historical data rollback when the heat pump power exceeds the limit (>130% of rated value) or the phase change material temperature changes abruptly (ΔT>5℃ / min). During system maintenance, the temperature drift error of the fiber optic temperature sensing array is calibrated monthly, and the fitness function weight coefficients of NSGA-II are updated to adapt to seasonal changes in thermal storage demand.

[0121] This invention addresses the problems of lagging thermal accumulation control, low utilization of phase change materials, and energy imbalance in traditional systems through the synergy of multi-scale thermal field analysis, dynamic regulation, and multi-objective optimization. Compared to existing technologies, it can accurately identify heat diffusion paths and dynamically allocate heat release priorities, significantly improving the thermal balance response speed and regulation accuracy, effectively suppressing abnormal temperature gradient fluctuations, and optimizing the matching degree of phase change material charge and discharge rates. This enhances the long-term stability and energy efficiency of the system, providing efficient and reliable technical support for cross-seasonal energy storage under complex geological conditions.

[0122] When the global thermal balance module of this invention is applied to energy piles for high-efficiency cross-seasonal energy storage, its hardware environment includes: a distributed fiber optic temperature measurement array (DTS) arranged along the axial and radial directions of the energy pile to monitor the underground temperature gradient; a heat flux density sensor installed at the heat exchange pipe wall and the ground interface to monitor the heat exchange flow rate and heat diffusion in real time; a heat pump system, including a variable frequency heat pump and a variable speed circulating pump, to regulate the flow rate of the heat exchange medium; a phase change energy storage material module embedded inside the energy pile to optimize the heat storage / release rate through controllable filling; and an intelligent control unit, including a nonlinear PID controller and a deep optimization algorithm calculation server, for dynamically adjusting the heat exchange strategy and receiving feedback data from a remote SCADA monitoring system to ensure the long-term stability of the energy storage system.

[0123] Two sets of high-efficiency inter-seasonal energy storage piles with identical structures were selected. Experiments were conducted using a global thermal balance module (Group A) and traditional fixed threshold heat transfer control (Group B), respectively, to verify the heat distribution optimization capability and long-term thermal storage stability of this invention. Group A employed a gradient heat release strategy, combined with nonlinear PID control, storage-to-release power ratio optimization, and phase change charge-discharge optimization, dynamically adjusting the heat transfer flow rate under different formation temperature variations. Group B used fixed threshold heat transfer control, adjusting the heat transfer flow rate according to a preset temperature threshold, without considering the dynamic changes in formation heat diffusion, and thus unable to optimize adjustment based on local heat accumulation. During the experiment, the same number of temperature sensors and heat flux density sensors were arranged axially and radially on both sets of energy piles to ensure the accuracy and comparability of data acquisition. The experimental environment was set to a typical winter heating season, with each experiment repeated five times to reduce the influence of environmental factors. At the start of the experiment, the same initial formation temperature was set, and both systems were operated under the same load conditions. The experimental record table is shown in Table 5.

[0124] Table 5. Experimental Record of Global Thermal Balance Module

[0125]

[0126] Experimental results show that Group A energy piles, which utilize the global thermal balance module, significantly outperform Group B, which employs fixed threshold heat exchange control, in terms of heat release uniformity, heat pump regulation response speed, storage-to-release power ratio matching accuracy, and formation temperature stability. Therefore, the global thermal balance module of this invention can dynamically optimize heat release under different formation conditions, improving the stability and heat exchange efficiency of the thermal storage system. Compared to traditional fixed threshold control methods, it offers superior thermal management capabilities and stronger long-term operational economy, providing an intelligent optimization strategy for energy piles with efficient cross-seasonal energy storage.

[0127] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result according to substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A high-efficiency, cross-seasonal energy storage energy pile; characterized in that: include: The dynamic coupling prediction module is used to perform multi-scale prediction of building heat load using the LA hybrid time series prediction model. When the building energy consumption monitoring system detects that the start-stop frequency of indoor temperature control equipment exceeds the threshold, the LA hybrid time series prediction model reconstructs the weight of historical load samples through a sliding time window mechanism, and simultaneously inputs meteorological forecast data into the three-dimensional heat conduction correction coefficient matrix to output the cross-seasonal heat storage demand prediction value. The formation temperature field reconstruction module is used to collect formation temperature field gradient data and axial and radial temperature gradients within the energy pile using a distributed fiber optic temperature measurement array. When the rate of change of the formation temperature gradient exceeds a preset critical value, the module reconstructs the three-dimensional unsteady temperature field based on the heat flux density sensor and temperature sensor data within the energy pile, calculates the heat diffusion loss rate, and corrects the flow velocity of the heat exchange tube and the filling density of the phase change material in real time. The geological parameter identification module, including the transient thermal response test unit and the pore water pressure monitoring array, is used to identify the geological type of sand, clay or bedrock through a support vector machine classifier when the borehole radar detects abnormal fluctuations in the dielectric constant of the formation, map the heat exchange parameters, and invert the equivalent thermal conductivity and the thermal storage density correction coefficient through the thermo-permeability coupling model. The adaptive optimization control module is used to adopt a convection-dominated heat transfer strategy when the groundwater flow velocity output by the geological parameter identification module exceeds a preset threshold or the thermal conductivity is lower than a preset threshold. The optimal control strategy is generated through a depth-deterministic strategy gradient algorithm, and the heat pump power and circulation pump frequency are adjusted in real time. At the same time, the trigger threshold of the phase change material is adjusted according to the real-time formation temperature distribution. The global thermal balance module is used to actively compensate for heat loss by adjusting the heat pump working fluid flow rate through a nonlinear PID controller when the temperature gradient of the thermal accumulation region exceeds a preset threshold detected by the formation temperature field reconstruction module. At the same time, it dynamically adjusts the heat charge and release rate of the phase change material according to the real-time heat storage and release power ratio.

2. The energy pile for high-efficiency cross-seasonal energy storage according to claim 1, characterized in that: The LA hybrid time series prediction model includes a hierarchical time series decomposition layer, a multimodal fusion layer, a dynamic weight reconstruction layer, a three-dimensional heat conduction correction layer, an anomaly frequency judgment layer, and a cross-seasonal coupling output layer; The hierarchical temporal decomposition layer is used to generate an eigenmode function set by superimposing adaptive white noise iteratively through an improved adaptive noise set empirical mode decomposition algorithm. The layer then filters out components whose sample entropy is lower than a preset threshold of 0.8 and whose energy percentage exceeds 75%, and outputs high-frequency noise, mid-frequency diurnal periodicity and low-frequency seasonal components. The multimodal fusion layer is used to calibrate the time series offset of meteorological and load data through a dynamic time warping algorithm. When the peak offset of the cross-correlation function between meteorological and load data exceeds a preset threshold, the time axis of meteorological data is stretched. Temperature, humidity, radiation parameters and load components are fused into a three-dimensional spatiotemporal tensor through a multi-head attention mechanism. The dynamic weight reconstruction layer is used to mark outliers based on the KL divergence distribution of samples within a sliding window using Bayesian optimization when the Mahalanobis distance exceeds a preset confidence interval threshold. For non-outlier samples, the maximum a posteriori probability weight matrix is ​​solved using Gaussian regression. The three-dimensional heat conduction correction layer is used to construct an unstructured formation heat conduction grid model using the finite element method. When the meteorological temperature change rate ΔT / Δt > 2℃ / h, the correction coefficient matrix is ​​iteratively solved using a transient heat conduction control formula. The transient heat conduction control formula is: In formula (1), ρ is the soil density, with units of kg / m³. 3 C p ρ represents the specific heat capacity of the soil, in J / (kg·K), T represents the ground temperature field, t represents the time variable, in seconds, k represents the thermal conductivity of the soil, in W / (m·K), and Q represents the soil thermal conductivity. solar This is the solar radiation heat source term, with units of W / m². 3 , This is the heat flux density correction matrix; The anomaly frequency decision layer is used to define the sequence of device start-up and shutdown events as observations based on the hidden Markov model. The implicit states include normal, transition and abnormal. If the proportion of the abnormal state Viterbi path is greater than the preset threshold, the sample pool is cleared and 30 days of historical data are reloaded. The cross-seasonal coupling output layer is used to input the predicted heat flux density into the pile strain formula: ε=aΔT+β+δ thermal / Δt (2) In formula (2), ε is the strain value of the pile body, a is the thermal expansion coefficient of the pile material, ΔT is the temperature change, and β is the thermal stress coupling coefficient, with units of MPa. -1 δ thermal The value is the thermal stress, in MPa. If ε > 0.2%, the Pareto optimal solution is obtained by sequential quadratic programming algorithm, and the thermal storage demand value is output and sent to the heat pump control unit.

3. The energy pile for high-efficiency cross-seasonal energy storage according to claim 1, characterized in that: The working method of the finite element inversion algorithm is as follows: Based on the coordinates of the anomalous regions marked by clustering, the target region is locally meshed using an unstructured mesh generation algorithm. Heat flux density sensor data and clustered partition temperature data are loaded, and a heat conduction forward model is used. A nonlinear least squares algorithm is employed to iteratively invert the formation's thermal properties, including thermal conductivity γ and heat capacity c, to reconstruct the three-dimensional unsteady temperature field and calculate the heat diffusion loss rate n. loss : In formula (3), γ is the dynamic correction value of the formation thermal conductivity; q sensor This represents heat flux density sensor data; v is the flow rate of the working fluid inside the heat exchanger tube. Let n be the temperature gradient; if n loss If the flow rate of the heat exchange tube exceeds the preset safety threshold, the flow rate of the heat exchange tube is dynamically adjusted by the PID controller, and the optimal filling density under the current temperature field is matched by a linear interpolation algorithm based on the latent heat temperature curve of the phase change material.

4. The energy pile for high-efficiency cross-seasonal energy storage according to claim 1, characterized in that: The transient thermal response testing unit applies a short-duration thermal pulse to the formation around the borehole using pulsed thermal flux excitation, and collects formation temperature response data through fiber Bragg grating sensing to calculate the transient thermal diffusion coefficient and identify the initial thermal properties of the formation. The pore water pressure monitoring array extracts the groundwater flow trend through a dynamic water pressure gradient analysis method and uses adaptive permeability regression (APR) to calculate soil permeability parameters, thereby obtaining the correction factor for the influence of groundwater on thermal diffusion.

5. The energy pile for high-efficiency cross-seasonal energy storage according to claim 1, characterized in that: In the geological parameter identification module, the support vector machine classifier operates as follows: First, the dielectric properties of the formation are projected into a high-dimensional feature space through kernel function mapping. Then, the Lagrange dual optimization method is used to maximize the data class interval, and soft-margin support vectors are used to enhance the model's adaptability to geological transition zones. During the classification process, the support vector machine classifier dynamically updates the support vectors using KKT conditional filtering to remove redundant data points. Subsequently, an adaptive weight adjustment mechanism optimizes the support vector weight distribution so that the decision boundary of the classifier between sand, clay, and bedrock categories conforms to the trend of formation thermal property changes. Finally, the support vector machine classifier uses a heat exchange parameter mapping mechanism to call the corresponding thermal diffusivity, heat capacity ratio, and permeability from a predefined thermal property parameter library based on the classification results.

6. The energy pile for high-efficiency cross-seasonal energy storage according to claim 1, characterized in that: In the formation thermal-seepage coupling inversion process, the aforementioned thermal-seepage coupling model first constructs unsteady thermal-seepage coupling equations based on the classified geological types, and uses a joint solution method of energy equation and Darcy's law to describe the dynamic interaction between underground fluids and the thermal field. Then, the relationship matrix between underground temperature field, pore water velocity and heat transfer capacity is nonlinearly discretized using the finite element multi-scale inversion method (FEMI). During the inversion process, the thermal-seepage coupling model uses the Poisson equation to constrain the PEC to correct the groundwater velocity. After the corrected output parameters are verified by sensitivity analysis, they are pushed to the control system to dynamically adjust the operation strategy of the thermal reservoir.

7. The energy pile for high-efficiency cross-seasonal energy storage according to claim 1, characterized in that: The adaptive optimization control module operates as follows: Based on the groundwater flow velocity and formation thermal conductivity anomaly signals from the geological parameter identification module, constraint conditions are established through a boundary constraint mechanism; subsequently, the heat pump power and circulation pump frequency are optimized and adjusted using a convection-dominated heat transfer strategy; during the control process, a multi-step time-series strategy is used to evaluate and calculate the cumulative return value under different heat transfer strategies, and the control parameters are corrected through a strategy gradient adaptive update method; next, the optimal trigger temperature of the phase change material is calculated based on a real-time formation temperature distribution prediction method, and the phase change thermal storage process is dynamically adjusted using a phase change trigger threshold dynamic adjustment mechanism; finally, the heat transfer parameters are iteratively adjusted through deep reinforcement learning feedback optimization to optimize the long-term thermal storage stability of the energy pile.

8. The energy pile for high-efficiency cross-seasonal energy storage according to claim 1, characterized in that: The working steps of the depth deterministic policy gradient algorithm include: Step 1: Based on the state-action value function, define the optimal time difference of the target policy and calculate the policy value. The formula is as follows: In formula (4), For the optimal time difference objective value, R t denoted as δ, where δ is the instantaneous benefit of the heat exchange control strategy at time t; Q' is the discount factor; Q' is the Q-value estimate of the target network; a' is the optimal strategy for the next step; and ω1 is the strategy smoothing factor. The action performed in the previous moment; θ Q’ For the target Q network parameters; S t+1 This indicates that action a is performed at the current time t. t Then, the changes in the environmental state at time t+1; Represents the policy network parameters θ u The gradient calculation value; Step 2: Optimize the policy network using the policy gradient update function, and calculate the gradient. The calculation formula is as follows: In formula (5), J(θ) u ) is the policy objective function; u(S|θ) u ) represents the policy network; ω2 is the gradient smoothing factor; Let S be the rate of change for policy updates; E represents the expected value of sampling all future states S and actions a. Step 3: Store historical state data through an experience replay mechanism, and optimize by introducing an entropy regularization term using an adaptive entropy regularization strategy to improve the exploration capability of the heat exchange strategy. The formula is: In formula (6), τ entropy ω is the entropy regularization loss; ω3 is the regularization coefficient; π(a) i |S) represents the policy probability distribution; N is the number of samples in the policy space; Step 4: Calculate the optimal heat transfer control parameters by combining the adaptive adjustment of the circulating pump frequency with the nonlinear mapping of the heat pump power. As the final execution strategy.

9. The energy pile for high-efficiency cross-seasonal energy storage according to claim 1, characterized in that: The working principle of the gradient heat release strategy is as follows: Based on the temperature gradient data of the formation temperature field reconstruction module, the spatial distribution of the heat accumulation region is calculated by the multi-scale thermal partitioning clustering method, and the heat diffusion path is identified by the gradient direction field fitting method GDFF to generate a dynamic heat release priority weight matrix; then, based on the dynamic feedback of the formation thermal diffusivity and the heat pump working fluid flow rate, the heat exchange flow rate is adjusted in real time by the nonlinear PID heat flow control method. The nonlinear PID heat flow control method suppresses the temperature fluctuation caused by unsteady heat transfer through adaptive disturbance compensation. Subsequently, the deviation between the current heat storage rate of the phase change material and the formation heat release demand was calculated using the heat storage-to-release power ratio adjustment method HSRP-AA, and the heat release rate of the phase change material was optimized by calling the phase change dynamic charge-discharge control method PCD-CDR to match the local temperature gradient change trend. Finally, based on the formation heat release rate, working fluid flow fluctuation and heat storage stability, the heat pump flow control strategy was adaptively updated using the multi-objective thermal balance iterative optimization method MOTB-IO.

Images