A photovoltaic-geothermal complementary coupling energy supply optimization scheduling method

Abstract

The application relates to the fields of data acquisition systems and intelligent scheduling control, and specifically discloses a photovoltaic-geothermal complementary coupling energy supply optimization scheduling method. The method constructs a thermodynamic model containing photovoltaic power generation, geothermal energy conversion, combined heat and power and energy storage devices, generates a Pareto optimal solution set offline based on annual meteorological and load data by using a multi-objective bionic optimization algorithm, trains a student neural network model through knowledge distillation, deploys the student neural network model on an edge end to realize real-time scheduling decision and dynamically adjust power distribution of each unit, and introduces an online performance evaluation and closed-loop feedback mechanism to support multi-time scale cooperation and adaptive switching in abnormal working conditions. Through the above technical scheme, the application improves the cooperation level of exergy efficiency, economy and environmental benefits of the system, and realizes second-level response and long-term operation stability.

Description

Technical Field

[0001] This invention belongs to the field of data acquisition systems and intelligent scheduling and control, and specifically relates to a photovoltaic-geothermal complementary coupling energy supply optimization scheduling method. Background Technology

[0002] As the global energy structure transitions towards a low-carbon model, multi-energy complementary distributed energy systems have become a key means to improve the utilization rate of renewable energy and ensure regional energy security. Photovoltaic power generation is clean but subject to random fluctuations, while geothermal energy boasts high stability and continuous power supply advantages. The coupled application of these two technologies can enhance system reliability and overall energy efficiency. In the construction of integrated energy microgrids, photovoltaic-geothermal complementary energy systems, through the deep coupling of cooling, heating, and electrical energy flows, provide an important pathway for achieving building energy conservation and regional carbon emission reduction.

[0003] Optimized scheduling is a core element in improving the operational efficiency of photovoltaic-geothermal complementary systems. This process involves the coordinated control of thermodynamic cycle equipment, photovoltaic arrays, and energy storage devices, aiming to achieve optimal energy allocation in both time and space by adjusting the distribution ratio of various energy flows. Its basic objective is to formulate the optimal operating strategy that meets the system's thermodynamic performance, economic costs, and environmental benefits based on meteorological forecast data and load demand.

[0004] Existing technologies often employ linear programming or mixed-integer linear programming methods, which struggle to accurately characterize the complex nonlinear thermodynamic characteristics of components such as organic Rankine cycles and absorption refrigeration systems. This leads to discrepancies between scheduling schemes and the actual operating state of the system. Photovoltaic output exhibits severe transient fluctuations due to environmental influences. Traditional optimization algorithms based on monitoring and data acquisition systems face problems such as excessive computational load, low search efficiency, and susceptibility to local optima when dealing with complex trade-offs involving multiple indicators including exergy efficiency, economic cost, and carbon emission reduction. These challenges make it difficult to meet the real-time dynamic scheduling requirements of electricity and heat loads. Furthermore, existing scheduling architectures lack knowledge accumulation and real-time decision feedback mechanisms, hindering the achievement of optimal system performance balance in dynamically changing environments. Summary of the Invention

[0005] The purpose of this invention is to provide a photovoltaic-geothermal complementary coupling energy supply optimization scheduling method, which can solve the problems in the background art mentioned above.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a photovoltaic-geothermal complementary coupling energy supply optimization scheduling method, comprising the following specific steps: Step 1: Construct a system thermodynamic model that includes a photovoltaic power generation unit, a geothermal energy conversion unit, a combined cooling, heating and power (CCHP) device, and an energy storage device. The thermodynamic model accurately characterizes the nonlinear energy conversion relationship in the organic Rankine cycle and absorption refrigeration process. Step 2: Based on the annual meteorological data and typical load curves, the thermodynamic model is optimized offline in the cloud server using a multi-objective biomimetic optimization algorithm to generate a Pareto optimal solution set covering three dimensions: exergy efficiency, levelized cost of energy, and carbon emission intensity. Step 3: Using a large number of mapping samples between input states and corresponding optimal scheduling strategies in the Pareto optimal solution set, train the student neural network model through knowledge distillation technology to enable it to learn the decision logic of complex optimization algorithms; Step 4: Deploy the trained deep neural network to the field monitoring and data acquisition system as a real-time scheduling decision module. Receive current weather forecast information, load demand, and equipment operating status, and output power allocation instructions for each energy unit in the next scheduling cycle. Step 5: Dynamically adjust the output of the photovoltaic inverter, the flow rate of the geothermal working fluid, the charging and releasing rate of the thermal storage tank, and the start-stop status of the chiller unit according to the power allocation command to achieve coordinated and optimized scheduling of multiple energy flows such as cold, heat, and electricity.

[0007] Preferably, the thermodynamic model constructed in step 1 includes a dynamic response mechanism to geothermal source temperature, ambient temperature, solar irradiance, and user-side cooling, heating, and power loads, and introduces an exergy loss coefficient to quantify the irreversibility of each energy conversion stage.

[0008] Preferably, the multi-objective biomimetic optimization algorithm used in step 2 is the Grey Wolf Optimization Algorithm or the Boundary Crossover Aggregation Multi-Objective Evolutionary Algorithm with Penalty Mechanism. Its population size, number of iterations and crossover mutation parameters are all preset according to the system degrees of freedom and constraints to ensure convergence to a high-quality Pareto front under limited computing resources.

[0009] Preferably, the offline optimization process in step 2 uses hours as the time granularity, covers 8760 time periods throughout the year, and takes into account equipment start-up and shutdown constraints, minimum running time limits, and energy storage capacity boundaries, generating no less than 100,000 sets of training samples for subsequent knowledge distillation.

[0010] Preferably, the knowledge distillation process in step 3 includes a joint training phase of the teacher model and the student model. The teacher model is the generator of the optimal solution of the original multi-objective optimization algorithm under a specific input, and the student model is a feedforward neural network with three hidden layers. Its activation function adopts the modified linear unit, and the output layer is normalized and mapped to the operating range of each controllable device.

[0011] Preferably, in step 3, the training process uses a dual loss function of mean squared error and classification accuracy, and adjusts the weights of the student model through the backpropagation algorithm until the deviation between its scheduling strategy on the validation set and the output of the teacher model is less than a preset threshold.

[0012] Preferably, the deep neural network deployed in step 4 is encapsulated as an independent executable module and integrated into the edge computing node of the on-site monitoring and data acquisition system. Its inference latency does not exceed 500 milliseconds, which meets the time requirements of intraday rolling scheduling.

[0013] Preferably, the meteorological forecast information received in step 4 includes the trend of solar irradiance change and the range of ambient temperature fluctuation in the next 2 hours, and the load demand data comes from the historical statistics and short-term forecast module of the building energy management system.

[0014] Preferably, the power allocation instructions in step 5 include the photovoltaic maximum power point tracking offset, the geothermal pump working fluid mass flow rate setting, the opening ratio of the inlet and outlet valves of the high-temperature thermal storage tank, and the speed instructions of the absorption chiller solution circulation pump. All instructions comply with the equipment safety operation specifications.

[0015] Preferably, the method further includes an online performance evaluation unit, which is used to calculate the exergy efficiency, unit energy cost and carbon emission equivalent under the current scheduling strategy in real time, and feed the evaluation results back to the cloud server for periodically updating the training dataset of the student model; The online performance evaluation unit performs an index calculation every 15 minutes. If the evaluation results deviate from the Pareto front by more than a predetermined distance for three consecutive times, the model retraining process is triggered to ensure that the scheduling strategy is always in the optimal performance range. The method supports multi-timescale collaborative scheduling, wherein the day-ahead plan is optimized offline in the cloud to generate a baseline curve, and the intraday rolling scheduling is dynamically corrected by the on-site neural network module based on the latest measured data, forming a closed-loop optimization architecture. The closed-loop optimization architecture is equipped with an abnormal operating condition identification mechanism. When events such as a sudden drop in photovoltaic output, abnormal geothermal well temperature, or sudden load change are detected, it automatically switches to a conservative scheduling mode to prioritize ensuring the continuity of power and heating for critical loads. In the conservative scheduling mode, the system prioritizes geothermal energy and energy storage devices to maintain the basic load, restricts the operation of unnecessary cooling equipment, and gradually restores to the normal optimized scheduling state after the operating conditions stabilize.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. By integrating thermodynamic modeling, multi-objective biomimetic optimization, and knowledge distillation techniques, a scheduling architecture combining offline high-precision optimization and online lightweight decision-making was constructed, overcoming the shortcomings of traditional linear programming methods in handling nonlinear thermodynamic processes. 2. This invention achieves a synergistic improvement in exergy efficiency, economic efficiency, and environmental benefits through a Pareto front-guided multi-objective trade-off mechanism, avoiding system performance imbalances caused by single-objective optimization. The knowledge distillation process distills the decision-making wisdom of complex optimization algorithms, enabling edge devices to have scheduling capabilities close to cloud-based optimization levels and reducing dependence on on-site computing resources. The closed-loop feedback mechanism ensures that the scheduling strategy can continuously and adaptively optimize as the system ages, weather patterns change, and load characteristics evolve, improving the long-term operational stability and overall energy efficiency of the photovoltaic-geothermal complementary system. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical solution architecture according to the present invention; Figure 2 This is a schematic diagram of the core principle framework of the teacher-student model strategy transfer based on knowledge distillation according to the present invention. Figure 3 This is a schematic diagram of the multi-level interaction relationship and data flow between cloud-based offline optimization and edge-based real-time scheduling according to the present invention; Figure 4 This is a flowchart illustrating the logical process of multi-objective thermodynamic modeling and Pareto optimal solution set generation according to the present invention. Figure 5 This is a closed-loop logic flowchart of real-time scheduling decision execution and performance adaptive evaluation feedback according to the present invention. Detailed Implementation

[0018] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0019] This embodiment provides a photovoltaic-geothermal complementary coupling energy supply optimization scheduling method. In implementing this method, the system integrates high-precision thermodynamic modeling technology, multi-objective biomimetic optimization algorithms, and a deep learning inference framework based on knowledge distillation, aiming to achieve efficient, economical, and low-carbon operation of complex energy systems in dynamic environments.

[0020] Step 1 involves constructing a system thermodynamic model comprising a photovoltaic power generation unit, a geothermal energy conversion unit, a combined cooling, heating, and power (CCHP) system, and an energy storage device. This thermodynamic model accurately characterizes the nonlinear energy conversion relationship in the organic Rankine cycle and absorption refrigeration process. During model construction, the photovoltaic power generation unit undergoes physical modeling, considering solar radiation intensity, photovoltaic module surface temperature, and module fill factor. The output power of the photovoltaic module is described as the product of incident radiation energy and photoelectric conversion efficiency, and is corrected in real-time using a temperature reduction factor based on changes in ambient temperature. The geothermal energy conversion unit primarily involves the extraction of thermal energy from geothermal fluids. In this process, a power consumption model for the geothermal downhole pump and a thermodynamic characterization of the wellhead fluid are established. The core of the CCHP system includes an organic Rankine cycle subsystem, which absorbs excess geothermal or photovoltaic heat energy through an evaporator, causing a phase change in the organic working fluid to generate high-pressure steam. This steam then enters an expander to drive a generator, producing electricity. The energy balance in this process is characterized in detail, including the enthalpy, pressure, and entropy of the working fluid at various state points. For absorption refrigeration processes, the model covers the thermodynamic balance equations of four core components: generator, absorber, condenser and evaporator. By describing the changes in mass concentration of lithium bromide solution or ammonia solution, the cycle ratio and the logarithmic mean temperature difference during the heat exchange process, the model accurately calculates the coefficient of performance (COP) of the system at different heat source temperatures.

[0021] Furthermore, the thermodynamic model constructed in step 1 includes a dynamic response mechanism to geothermal source temperature, ambient temperature, solar irradiance, and user-side heating, cooling, and electrical loads, and introduces an exergy loss coefficient to quantify the irreversibility of each energy conversion stage. The exergy analysis focuses not only on the quantity of energy but also on its quality. The model identifies loss points within the system by calculating the difference between input and output exergy. For example, in an organic Rankine cycle evaporator, the exergy loss due to the heat transfer temperature difference is defined as the product of the heat transfer rate and the reciprocal difference between the ambient temperature and the heat source temperature. For energy storage devices, including batteries and thermal storage tanks, the model introduces charge / discharge depth constraints, self-discharge loss rate, and a dynamic equation for the state of charge evolving over time. The modeling of thermal storage tanks also considers the stratification effect, i.e., simulating the water temperature distribution at different heights within the tank using a multi-node method to improve the accuracy of heat extraction and storage prediction.

[0022] Step 2 involves using a multi-objective biomimetic optimization algorithm on a cloud server to offline optimize the thermodynamic model based on annual meteorological data and typical load curves. This generates a Pareto optimal solution set covering three dimensions: exergy efficiency, levelized cost of energy (LCOE), and carbon emission intensity. The cloud server, as the computational core, acquires long-sequence historical meteorological data for specific geographical locations, including hourly measurements of total horizontal radiation, direct radiation, diffuse radiation, and dry-bulb temperature. The typical load curves cover the cooling, heating, and electricity demand patterns of residential, commercial, or industrial users throughout the four seasons. The optimization process aims to find an optimal set of equipment capacity configuration and scheduling baseline parameters that maximizes exergy efficiency throughout the system's lifespan and minimizes the LCOE per unit of energy output and carbon dioxide emissions per unit of energy supplied.

[0023] In step 2, the multi-objective biomimetic optimization algorithm used is either the gray wolf optimization algorithm or a boundary crossover and aggregation multi-objective evolutionary algorithm with a penalty mechanism. The population size, number of iterations, and crossover mutation parameters are all pre-set according to the system's degrees of freedom and constraints to ensure convergence to a high-quality Pareto front under limited computational resources. Taking the gray wolf optimization algorithm as an example, the system simulates the social hierarchy and hunting behavior of a gray wolf population, dividing the potential optimal solutions in the solution space into alpha wolves, secondary wolves, follower wolves, and bottom wolves. By calculating the distance between the current wolf and the alpha, secondary, and follower wolves, and utilizing the synergistic effect of convergence factors and random weights, the wolf's position information is continuously updated, achieving global search and local refinement of the nonlinear solution space. To handle hard constraints in system operation, such as the upper and lower limits of generator output and the charging and discharging rate limits of energy storage devices, the algorithm introduces a dynamic penalty function. When an individual violates a constraint, its objective function value will be given a large penalty increment according to the degree of violation, guiding the population to converge towards the feasible region during the evolutionary process.

[0024] Furthermore, in step 2, the offline optimization process uses hours as the time granularity, covering 8760 time periods throughout the year, and considers equipment start-up and shutdown constraints, minimum operating time limits, and energy storage capacity boundaries, generating no fewer than 100,000 training samples for subsequent knowledge distillation. Each training sample contains an environmental state vector at that moment, such as meteorological parameters, load demand, current energy storage capacity, and the corresponding optimal control vector given by the optimization algorithm, such as the power setpoints, valve opening degrees, and on / off states of each unit. By performing large-scale parallel simulations in the cloud, the system constructs a multi-dimensional state-decision mapping space, which provides a complete knowledge base for subsequent neural network training.

[0025] Step 3 involves using a large number of mapping samples between input states and corresponding optimal scheduling strategies in the Pareto optimal solution set to train a student neural network model using knowledge distillation technology. This allows the model to learn the decision-making logic of complex optimization algorithms. The core of knowledge distillation lies in using the soft labels output by a high-performance but computationally intensive "teacher model" to guide a simple and fast-running "student model." In this embodiment, the teacher model is the multi-objective biomimetic optimization algorithm running in the aforementioned steps.

[0026] In step 3, the knowledge distillation process includes a joint training phase for the teacher and student models. The teacher model is the generator of the optimal solution for the original multi-objective optimization algorithm under specific inputs, while the student model is a feedforward neural network with three hidden layers. Its activation function uses modified linear units (MRUs), and the output layer is normalized and mapped to the operating range of each controllable device. The design of the student neural network fully considers the inference capabilities of the hardware platform, and the number of neurons in its hidden layers is optimized through pruning to balance prediction accuracy and computational cost. The application of modified linear units can alleviate the gradient vanishing problem during deep network training and improve the convergence speed of the model.

[0027] Furthermore, in step 3, the training process employs a dual loss function of mean squared error and classification accuracy. The weights of the student model are adjusted using a backpropagation algorithm until the deviation between its scheduling strategy on the validation set and the teacher model's output is less than a preset threshold. The mean squared error loss function measures the Euclidean distance between the continuous control variables output by the student model, such as power values, and the values ​​suggested by the teacher model; the classification accuracy loss function evaluates discrete control logic, such as equipment start-up and shutdown decisions. By introducing a temperature coefficient to smooth the output distribution, the student model can learn the underlying logical connections behind the teacher model's decisions, rather than simply performing numerical fitting. The training process is completed on a cloud-based GPU-accelerated platform, and through multiple rounds of iterative optimization, the student model is ensured to have good robustness under various extreme weather and fluctuating load scenarios.

[0028] Step 4 involves deploying the trained deep neural network to the field monitoring and data acquisition system as a real-time scheduling decision module. This module receives current weather forecasts, load demands, and equipment operating status, and outputs power allocation instructions for each energy unit in the next scheduling cycle. The field monitoring and data acquisition system, also known as the SCADA system, is responsible for low-level communication with various sensors and actuators.

[0029] In step 4, the deployed deep neural network is encapsulated as a standalone executable module and integrated into the edge computing node of the on-site monitoring and data acquisition system. Its inference latency does not exceed 500 milliseconds, meeting the time requirements of intraday rolling scheduling. The edge computing node has a high-performance embedded processor capable of parsing real-time data packets from the bus or wireless network. The limitation on inference latency ensures that the scheduling module can respond quickly when photovoltaic output changes abruptly due to cloud cover.

[0030] Furthermore, in step 4, the received meteorological forecast information includes the trend of solar irradiance change and the range of ambient temperature fluctuations over the next two hours. Load demand data is sourced from the historical statistics and short-term forecasting modules of the building energy management system. The system utilizes short-term forecasting algorithms, such as long short-term memory networks or combined forecasting models, combined with minute-level data measured on-site, to provide accurate feedforward input to the deep neural network scheduling module. Equipment operating status includes real-time telemetry parameters such as the current voltage and charge of the energy storage device, the temperature gradient of the thermal storage tank, and the cumulative operating time of the generator set.

[0031] Step 5 involves dynamically adjusting the output of the photovoltaic inverter, the flow rate of the geothermal working fluid, the charging and discharging rate of the thermal storage tank, and the start-stop status of the chiller unit according to the power allocation command, thereby achieving coordinated and optimized scheduling of multiple energy flows, including cold, heat, and electricity. This process translates logical commands into physical execution actions through industrial control protocols such as Modbus, Profinet, or the IEC61850 protocol specific to power systems.

[0032] In step 5, the power allocation command includes the photovoltaic maximum power point tracking offset, the geothermal pump working fluid mass flow rate setpoint, the opening ratio of the inlet and outlet valves of the high-temperature thermal storage tank, and the speed command of the absorption chiller solution circulation pump. All commands comply with the equipment safety operation specifications. For example, when photovoltaic output is excessive and grid acceptance capacity is limited, the command may guide the inverter to operate away from the maximum power point or increase the charging power of the thermal storage tank. When the heat load surges, the dispatch command will increase the frequency of the inverter driving the geothermal pump, increase the flow rate of the circulating working fluid, and synchronously adjust the heat source side valves of the absorption chiller to ensure priority allocation of heat energy.

[0033] The method also includes an online performance evaluation unit, used to calculate the exergy efficiency, unit energy cost, and carbon emission equivalent under the current scheduling strategy in real time, and feeds the evaluation results back to the cloud server for periodically updating the training dataset of the student model. The online performance evaluation unit performs an index calculation every 15 minutes. If three consecutive evaluation results deviate from the Pareto front by more than a predetermined distance, a model retraining process is triggered to ensure that the scheduling strategy is always within the optimal performance range. This closed-loop feedback mechanism enables the system to have self-learning and self-evolution capabilities, and can cope with system characteristic shifts caused by equipment aging, fouling, or long-term climate change.

[0034] The method supports multi-timescale collaborative scheduling. The day-ahead planning is optimized offline in the cloud to generate a baseline curve, while the intraday rolling scheduling is dynamically corrected by the on-site neural network module based on the latest measured data, forming a closed-loop optimization architecture. During the day-ahead phase, the system generates a coarse scheduling blueprint for the next 24 hours based on weather forecasts; during the intraday phase, a lightweight neural network fine-tunes the baseline curve at minute or second-level frequencies to offset the impact of prediction errors.

[0035] The closed-loop optimization architecture incorporates an abnormal operating condition identification mechanism. When events such as a sudden drop in photovoltaic output, abnormal geothermal well temperature, or sudden load changes are detected, the system automatically switches to a conservative scheduling mode, prioritizing the continuity of power and heating for critical loads. In conservative scheduling mode, the system prioritizes the use of geothermal energy and energy storage devices to maintain the base load, restricting the operation of unnecessary cooling equipment. Once the operating conditions stabilize, the system gradually returns to the normal optimized scheduling state. The anomaly identification algorithm monitors the magnitude of residuals or utilizes statistical process control-based methods to determine in real time whether the current operating conditions exceed the training coverage of the neural network, providing safety redundancy in extreme cases.

[0036] Example 2: Based on Example 1, this example provides a more detailed description of the specific physical parameters and logical implementation of the thermodynamic model. For a photovoltaic-geothermal complementary coupled energy supply system, the core lies in how to maximize benefits through energy gradient utilization.

[0037] The photovoltaic power generation unit in the system's thermodynamic model not only considers electrical energy output but also couples with a photovoltaic waste heat recovery module. By laying heat exchange pipes on the back of the photovoltaic panels, the cooling medium removes the heat from the panels, improving photoelectric conversion efficiency while generating medium-to-low temperature hot water. This hot water serves as an auxiliary heat source for the geothermal energy conversion unit and enters the preheater of the organic Rankine cycle. In the model, the temperature of the photovoltaic modules is described as a nonlinear dynamic equilibrium process influenced by ambient wind speed, ambient temperature, solar radiation, and cooling medium flow rate. As the cooling medium flow rate increases, the panel temperature decreases, and the degree of improvement in photoelectric conversion efficiency is quantified using linear regression coefficients.

[0038] The geothermal energy conversion unit involves the development of deep geothermal resources, and the model provides a detailed characterization of the thermodynamic coupling between geothermal production wells and reinjection wells. Utilizing the principles of porous media fluid dynamics, the heat transport process within the geothermal reservoir is described. Key parameters in the model include the dynamic viscosity, density, and heat capacity of the geothermal fluid, all of which are set as functions of temperature. When the geothermal fluid enters the surface heat exchanger, the heat transfer between it and the organic Rankine cycle working fluid follows the fundamental law of heat transfer, namely the product of the heat transfer coefficient, the heat transfer area, and the effective average temperature difference. To improve accuracy, the heat exchanger is subdivided into dozens of micro-segments, each of which undergoes individual energy and mass conservation calculations.

[0039] In the offline optimization stage of step 2, the objective function is set using a weighted sum method or a constraint method. The exergy efficiency is calculated by dividing the total useful work output by the system (the sum of the effective energy of electrical energy and thermal energy) by the total exergy input by the system (the sum of solar radiation exergy and geothermal fluid exergy). The levelized cost of energy (LCOE) integrates the initial investment cost, operation and maintenance costs, equipment residual value, and discount rate, allocating them to the total power generation over the entire life cycle. Carbon intensity is calculated by summing the average carbon intensity coefficient of electricity production and the emissions generated during geothermal development.

[0040] In the execution flow of the multi-objective biomimetic optimization algorithm, the system performs non-dominated sorting in each generation of evolution. This involves comparing the performance of different solutions across the three objective dimensions to determine which solutions are not completely superior to others. These non-dominated solutions are stored in an external archive, forming the Pareto front. To ensure the uniformity of solution distribution, the algorithm introduces a crowding distance index, prioritizing the retention of individuals located in sparse regions of the solution space. In the boundary cross-aggregation multi-objective evolutionary algorithm, a set of uniformly distributed reference direction vectors is defined to decompose the multi-objective problem into multiple sub-problems for collaborative optimization. Each individual is primarily responsible for finding the optimal boundary in a specific direction.

[0041] In step 3, feature engineering was performed on the input vector of the deep neural network. In addition to the original physical quantities, features representing the rate of change of energy supply and demand imbalance, and the number of charge-discharge cycles of the energy storage device, were introduced. These higher-order features help the network sensitively capture the dynamic characteristics of the system. During training, the system adopted a learning rate decay strategy, that is, as the number of iterations increases, the step size of weight updates is reduced to ensure that the model can accurately lock onto the vicinity of the global minimum of the loss function.

[0042] Each hidden layer of the student neural network is equipped with a layer normalization operation. Layer normalization forces the output distribution of each neuron in the layer to a standard distribution with a mean of 0 and a variance of 1, accelerating network training convergence and improving generalization ability. In the output layer, a multi-head output structure is used to accommodate control variables with different physical meanings. For example, one head is responsible for outputting continuous power setpoints using a linear activation function; another head is responsible for the logical judgment of device start / stop, using a logistic regression function to output probability values, and then determining the final state through a set threshold.

[0043] In step 4, the edge computing nodes of the SCADA system maintain a heartbeat connection with the cloud server. The cloud server periodically pushes updated weight parameters to the edge nodes. The deployment process adopts a canary release mode, that is, the scheduling effect of the new model is first tested on a small scale, and after comparative experiments confirm that its performance is better than the old model, a full replacement is then carried out. The task scheduler of the edge nodes is configured with high priority to ensure that the neural network inference task receives sufficient computing resources in the operating system.

[0044] In the instruction execution phase of step 5, to prevent excessively frequent control commands from causing equipment wear, the system introduces dead time and minimum fluctuation threshold logic. When the power fluctuation command output by the neural network is less than 2% of the equipment's rated power, the command is ignored to maintain system operational stability. The geothermal pump's working fluid flow regulation employs a variable frequency control strategy, changing the synchronous speed of the motor stator winding by adjusting its power supply frequency. The opening ratio of the inlet and outlet valves of the heat storage tank is smoothly adjusted according to the distribution of the water temperature gradient to avoid water hammer or thermal shock.

[0045] Example 3: This example focuses on the closed-loop feedback and adaptive optimization mechanism between steps 4 and 5, and how to handle the decoupling and coordinated control of multiple energy flows.

[0046] During real-time scheduling, the online performance evaluation unit not only monitors static metrics but also analyzes the system's dynamic response characteristics. For example, the system calculates the time required for the system to recover to steady state after a photovoltaic fluctuation triggers a scheduling command; this is the settling time. If the settling time is too long, it indicates that the control gain setting of the current student model may no longer be suitable for the existing hardware conditions, and the evaluation unit will record this deviation.

[0047] The specific performance calculation logic is as follows: The system collects flow, temperature, pressure, current, and voltage signals from each energy branch in real time. These signals are low-pass filtered to remove high-frequency noise interference. The exergy efficiency is updated every minute. At the end of the calculation cycle (15 minutes), the system calculates the average efficiency, total operating cost, and total emission reduction for that cycle. If these data deviate from the offline-generated Pareto front envelope, the evaluation system will determine that the current model has entered a sub-optimal state.

[0048] The process of triggering model retraining includes the following sub-steps: the edge node packages the current running state vector and the actual generated optimal decision (obtained through local small-scale search or rule correction) into an incremental dataset and uploads it to the cloud; the cloud server merges this new data with the original training set, assigns higher weights to the new data, and uses transfer learning techniques to incrementally fine-tune the neural network; the fine-tuned model is verified for security in the simulator, and after passing the verification, it is redeployed to the edge node.

[0049] In terms of multi-energy flow collaborative optimization, the method in this embodiment introduces an energy hub model to represent the three energy forms—cooling, heating, and electricity—within a unified mathematical framework. The system defines an energy conversion matrix, where each element represents the conversion efficiency from one energy form to another. Due to the strong thermodynamic coupling between the organic Rankine cycle and the absorption chiller (e.g., they share a heat source, or the heat exhaust of one can serve as a heat source for the other), the scheduling neural network learns this deep coupling constraint. When the cooling load demand is high and the electrical load demand is low, the network intelligently adjusts the proportion of heat medium flowing to the chiller, reducing the load on the expander and achieving optimal energy allocation in both time and space.

[0050] For the specific implementation of the abnormal operating condition identification mechanism, this embodiment adopts an anomaly detection logic based on reconstruction error. During training, the deep neural network not only learns the scheduling strategy, but also learns the data distribution characteristics under normal operating conditions through parallel autoencoder branches.

[0051] The parallel autoencoder branch adopts the same input layer structure as the student neural network, containing two symmetrical sub-modules: an encoder and a decoder. The encoder consists of two fully connected layers: the first layer has half the number of neurons as the input dimension, and the second layer has one-quarter the number of neurons as the input dimension, compressing the high-dimensional sensor vector into a low-dimensional latent space feature vector. The decoder consists of two fully connected layers symmetrical to the encoder, reconstructing the latent space feature vector into a vector of the original input dimension. The training process of the autoencoder is synchronized with that of the student neural network, using an independent mean squared error loss function. Only historical data under normal system operation is used as training samples, excluding any abnormal operating condition data. During training, by minimizing the mean squared error between the input vector and the reconstructed vector, the autoencoder learns the inherent correlation patterns between sensor data under normal operating conditions. After training, the weight parameters of the autoencoder branch are frozen, and the autoencoder weights are only updated synchronously during incremental retraining of the student neural network, ensuring consistency between the anomaly detection logic and the scheduling decision logic.

[0052] During real-time operation, if the reconstruction error of the input sensor vector after passing through the autoencoder exceeds the preset statistical upper limit, it indicates that the system has encountered a rare fault condition or extreme weather condition. Control is immediately transferred to the preset expert rule base.

[0053] The reconstruction error is quantified using a weighted mean square error, and its calculation formula is as follows: ; The reconstruction error at the current moment. Let the dimension be the input sensor vector. Let be the measured value of the i-th sensor. Let be the reconstructed value of the i-th sensor by the automatic encoder. is the weighting coefficient for the i-th sensor. The weighting coefficient is set according to the importance and measurement accuracy of the sensors. The weighting coefficient of key sensors such as photovoltaic output sensor, geothermal fluid inlet and outlet temperature sensor, energy storage state of charge sensor, and cold, heat and electricity load sensor is set to 2.0, while the weighting coefficient of non-key sensors such as auxiliary equipment operation status sensor and ambient humidity sensor is set to 1.0.

[0054] The preset statistical upper limit is dynamically determined using kernel density estimation combined with the 3σ criterion. The specific steps are as follows: In the initial stage of system operation, minute-level sensor data were collected during 30 consecutive days of normal operation, totaling 43,200 valid samples, and obvious sensor noise data was removed. The above samples are sequentially input into the trained autoencoder, and the reconstruction error corresponding to each group of samples is calculated to form an initial reconstruction error sample set. The kernel density of the reconstruction error sample set is estimated using the Gaussian kernel function to obtain the probability density distribution curve of the reconstruction error; Calculate the mean of the reconstruction error sample set. and standard deviation Set the initial value of the statistical upper limit to 0. This value can theoretically cover 99.7% of normally operating samples; The initial statistical upper limit was verified using known simulated abnormal operating conditions (such as a sudden drop of 30% in photovoltaic output, a 5°C drop in geothermal well temperature, and a sudden increase of 20% in load) to ensure that it could accurately identify all preset anomalies within 10 seconds. During system operation, the reconstruction error sample set is automatically updated monthly (retaining the normal operation data of the most recent 30 days) and the statistical upper limit is recalculated to adapt to the slow changes in system characteristics caused by equipment aging, fouling, etc.

[0055] The control strategy under conservative scheduling mode is as follows: The system disconnects non-critical auxiliary loads and maintains a minimum heat supply by locking the geothermal pump at a safe baseline speed. The energy storage device is set to backup power mode at this time, and deep cycle discharge is strictly prohibited to cope with possible total power outages. Only after the sensor readings return to the normal range and stabilize for more than 10 minutes will the system attempt to reactivate the neural network scheduling module and slowly recover to the optimized state.

[0056] This embodiment also provides a data cleaning logic based on cloud-edge collaboration. Since field sensors may drift or malfunction, edge nodes perform consistency checks before inputting data into the neural network. For example, the temperature of the geothermal fluid entering the heat exchanger must be higher than the outflow temperature, and the thermal balance deviation should be within a reasonable range. If a sensor reading with a logical error is detected, the system uses the correlation information of neighboring sensors, combined with a Kalman filter algorithm, to reconstruct and replace the abnormal reading, ensuring that the input source for scheduling decisions is reliable and continuous. This multi-level data verification and fault-tolerant design improves the operational reliability of the photovoltaic-geothermal complementary coupling energy supply system in unattended environments.

[0057] At the software architecture level, this approach employs containerized deployment technology. The deep neural network scheduling module is encapsulated within a lightweight container, enabling it to run cross-platform on edge computing hardware of different brands or models. A resource monitoring program within the container monitors memory and processor usage in real time. If resource exhaustion is detected, the system automatically restarts the scheduling service or reduces the frequency of data collection to ensure uninterrupted core scheduling logic.

[0058] This highly integrated scheduling framework not only solves the computational challenges posed by nonlinear thermodynamic modeling, but also brings complex cloud-based intelligence down to resource-constrained industrial sites through knowledge distillation technology. Experimental data shows that after adopting the method described in this embodiment, the system's annual comprehensive exergy efficiency is improved, the computational latency for intraday real-time scheduling is reduced from the traditional several seconds to the order of hundreds of milliseconds, the impact of distributed photovoltaic access on the local energy network is mitigated, and an efficient and stable balance of multiple energy flows is achieved.

[0059] At the specific hardware execution layer, the photovoltaic inverter employs a high-precision stepping frequency when executing maximum power point tracking offset commands. When the command requires a reduction in output to protect the grid, the inverter adjusts its internal pulse width modulation duty cycle to change the input impedance, shifting the operating voltage of the photovoltaic array towards open-circuit voltage. This process is limited within the permissible operating range of the photovoltaic cells, preventing overvoltage damage.

[0060] The circulating pumps on the geothermal fluid side are equipped with pressure closed-loop control. When a dispatch command requires a change in mass flow rate, the pump control logic automatically calculates the required head and monitors the pressure gradient in the pipeline network to ensure that cavitation does not occur during flow regulation. For absorption refrigeration systems, the speed regulation of the solution circulating pump is synchronized with the heating capacity of the generator to prevent excessively high lithium bromide solution concentration from causing crystallization in the heat exchanger. This fine-grained control at the physical level, combined with the intelligent decision-making of the upper-level neural network, constitutes a complete and highly robust energy dispatching system.

[0061] The method described in this embodiment is not only applicable to newly built photovoltaic-geothermal coupled energy supply systems, but can also be used to upgrade the energy efficiency of existing facilities through software upgrades of existing SCADA systems. By integrating neural network inference modules into existing industrial gateways and establishing communication links with the cloud, legacy energy systems can acquire predictive and adaptive optimization scheduling capabilities, providing solid technical support for the construction of large-scale regional energy internet.

[0062] Considering the real-time calculation of environmental benefits, the system simultaneously estimates the carbon emission reduction contribution corresponding to each set of dispatch instructions. This data is encrypted and stored in a local database and periodically aggregated and uploaded to the energy management platform. This not only provides data support for enterprises to achieve their carbon neutrality goals but also provides reliable original evidence for subsequent participation in carbon emission trading. By using technical means to bind physical energy flows with environmental value at the legal and policy level, this invention demonstrates its forward-looking and comprehensive advantages in modern energy management systems.

[0063] The solution in this embodiment ensures that the photovoltaic-geothermal complementary energy supply system can operate near the optimal performance solution set under any complex operating conditions through multi-level logical nesting and comprehensive physical constraint verification, truly achieving the optimal trade-off in energy utilization.

[0064] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A photovoltaic-geothermal complementary coupling energy supply optimization scheduling method, characterized in that, Includes the following steps: Step 1: Construct a system thermodynamic model that includes a photovoltaic power generation unit, a geothermal energy conversion unit, a combined cooling, heating and power (CCHP) device, and an energy storage device. The thermodynamic model characterizes the energy conversion relationship between the organic Rankine cycle and the absorption refrigeration process. Step 2: Based on meteorological data and typical load curves, a multi-objective biomimetic optimization algorithm is used in the cloud server to perform offline optimization of the thermodynamic model, generating a Pareto optimal solution set covering three dimensions: exergy efficiency, levelized cost of energy, and carbon emission intensity. Step 3: Using the mapping samples of input states and corresponding optimal scheduling strategies in the Pareto optimal solution set, train the student neural network model through knowledge distillation technology to enable it to learn the decision logic of multi-objective biomimetic optimization algorithms; Step 4: Deploy the trained student neural network model to the on-site monitoring and data acquisition system as a real-time scheduling decision module. Receive current weather forecast information, load demand, and equipment operating status, and output power allocation instructions for each energy unit in the next scheduling cycle. Step 5: Adjust the output of the photovoltaic inverter, the flow rate of the geothermal working fluid, the charging and releasing rate of the thermal storage tank, and the start-stop status of the chiller unit according to the power distribution command.

2. The photovoltaic-geothermal complementary coupling energy supply optimization scheduling method according to claim 1, characterized in that, The thermodynamic model constructed in step 1 includes: A photovoltaic power generation physical model is provided, which considers solar radiation intensity, photovoltaic module surface temperature and module fill factor. The output power of the photovoltaic module is defined as the product of incident radiation energy and photoelectric conversion efficiency, and is corrected by temperature reduction factor according to changes in ambient temperature. The geothermal energy conversion model includes a power consumption sub-model of the geothermal downhole pump and a thermodynamic characterization sub-model of the wellhead fluid, wherein the heat transfer between the geothermal fluid and the organic Rankine cycle working fluid is determined by the product of the heat transfer coefficient, the heat transfer area and the effective average temperature difference. An energy balance model is used to characterize the enthalpy, pressure, and entropy of the working fluid at each state point in an organic Rankine cycle, and to describe the changes in the mass concentration of lithium bromide solution or ammonia solution, the cycle ratio, and the logarithmic mean temperature difference during the heat exchange process in the absorption refrigeration process. A dynamic response mechanism is introduced, which introduces an exergy loss coefficient to quantify the irreversibility of the energy conversion process, wherein the exergy loss is defined as the product of the heat transfer and the reciprocal difference between the ambient temperature and the heat source temperature.

3. The photovoltaic-geothermal complementary coupling energy supply optimization scheduling method according to claim 1, characterized in that, The multi-objective biomimetic optimization algorithm used in step 2 is either the Grey Wolf Optimization Algorithm or the Boundary Cross-Aggregation Multi-Objective Evolutionary Algorithm with a Penalty Mechanism. The specific optimization process includes: The social hierarchy and hunting behavior of gray wolf populations are simulated. The potential optimal solutions in the solution space are divided into alpha wolves, secondary wolves, follower wolves and bottom wolves. By calculating the distance between the current wolf and the alpha wolf, secondary wolves and follower wolves, the position information of the wolves is updated by the synergistic effect of convergence factor and random weight, so as to realize the global search of the nonlinear solution space. In each generation of evolution, non-dominated sorting is performed. By comparing the performance of different solutions in the dimensions of exergy efficiency, levelized energy cost and carbon emission intensity, non-dominated solutions are identified and stored in an external archive to form a Pareto front. A crowding distance index is introduced to ensure the uniformity of solution distribution, and individuals located in sparse regions of the solution space are preferentially retained; In the boundary cross-aggregation multi-objective evolutionary algorithm, by defining a set of uniformly distributed reference direction vectors, the multi-objective problem is decomposed into multiple sub-problems for collaborative optimization, so that each individual finds the optimal boundary in the direction.

4. The photovoltaic-geothermal complementary coupling energy supply optimization scheduling method according to claim 1, characterized in that, The offline optimization process in step 2 includes the following in the time dimension: It uses hours as the time granularity, covers all time periods throughout the year, and introduces equipment start-up and shutdown constraints, minimum operating time limits, and energy storage capacity boundaries; Generate a preset number of valid training samples. Each training sample contains an environmental state vector and an optimal control vector. The environmental state vector consists of meteorological parameters, load demand and current energy storage margin, while the optimal control vector consists of the power setpoint, valve opening and switching status of each unit. By performing parallel simulations in the cloud, a multi-dimensional state and decision mapping space is constructed.

5. The photovoltaic-geothermal complementary coupling energy supply optimization scheduling method according to claim 1, characterized in that, The knowledge distillation process and the construction of the student neural network model in step 3 include: A joint training mode of teacher model and student model is adopted, in which the teacher model is the generator of the optimal solution of the original multi-objective optimization algorithm under the input, and the student model is a feedforward neural network with three hidden layers. The student neural network model is pruned and optimized to balance prediction accuracy and computational cost. Each neuron in its hidden layer is equipped with a layer normalization operation to transform the output distribution of each neuron into a standard distribution with a mean of zero and a variance of one. The activation function of the student neural network model is a modified linear unit, and its output layer adopts a multi-head output structure, with one head responsible for outputting the continuous power setpoint and the other head responsible for the logical judgment of device start-up and shutdown.

6. The photovoltaic-geothermal complementary coupling energy supply optimization scheduling method according to claim 1, characterized in that, The loss function setting and weight adjustment in step 3 of the training process include: We employ a dual loss function: mean squared error and classification accuracy. The mean squared error loss function measures the Euclidean distance between the continuous control variables output by the student model and the suggested values ​​by the teacher model, while the classification accuracy loss function evaluates the discrete control logic. By introducing a temperature coefficient to smooth the output distribution, the student model learns the underlying logical connections behind the teacher model's decisions. The backpropagation algorithm is used to adjust the weights of the student model, and a learning rate decay strategy is used to gradually reduce the step size of weight updates as the number of iterations increases, until the deviation between the scheduling strategy of the student model on the validation set and the output of the teacher model is less than a preset threshold.

7. The photovoltaic-geothermal complementary coupling energy supply optimization scheduling method according to claim 1, characterized in that, The deployment of the student neural network model and edge computing processing in step 4 include: The trained student neural network model is packaged into an independent executable module and integrated into the edge computing node of the on-site monitoring and data acquisition system, and the embedded processor parses the data packets in the communication network in real time. Set an upper limit for inference delay to ensure that the scheduling module responds within a preset millisecond time when photovoltaic output changes abruptly; Establish a heartbeat connection between the edge computing nodes and the cloud server. The cloud server periodically pushes updated weight parameters to the edge nodes and uses a canary release mode to replace the entire model.

8. The photovoltaic-geothermal complementary coupling energy supply optimization scheduling method according to claim 1, characterized in that, The information source and preprocessing process received in step 4 include: Meteorological forecast information includes the trend of solar irradiance change and the range of ambient temperature fluctuation within the future preset period. Load demand data comes from the historical statistics and short-term forecast modules of the building energy management system. Short-term prediction algorithms combined with on-site measured data are used to provide feedforward input for student neural network models; The equipment operating status includes the current voltage and power of the energy storage device, the temperature gradient of the thermal storage tank, and the cumulative operating time of the generator set; Before inputting data into the student neural network model, a consistency check is performed. If a sensor reading with a logical error is detected, the abnormal reading is reconstructed using the correlation information of adjacent sensors combined with the Kalman filter algorithm.

9. The photovoltaic-geothermal complementary coupling energy supply optimization scheduling method according to claim 1, characterized in that, The execution of power allocation commands and device control in step 5 include: The logic instructions are translated into physical execution actions through industrial control protocols. The power allocation instructions include the photovoltaic maximum power point tracking offset, the geothermal pump working fluid mass flow rate setpoint, the opening ratio of the inlet and outlet valves of the high-temperature thermal storage tank, and the speed instructions of the absorption chiller solution circulation pump. The logic of dead time and minimum change threshold is introduced. When the output power change command is less than the preset ratio of the device's rated power, the power change command is ignored to maintain system stability. The working fluid flow rate regulation of the geothermal pump adopts a variable frequency control strategy, which changes the synchronous speed by adjusting the power supply frequency of the stator winding of the motor; the speed of the solution circulation pump in the absorption refrigeration system maintains a synchronous proportional relationship with the heating capacity of the generator.

10. The photovoltaic-geothermal complementary coupling energy supply optimization scheduling method according to claim 1, characterized in that, The method also includes online performance evaluation, closed-loop optimization, and anomaly handling mechanisms. The online performance evaluation unit periodically performs index calculations, and calculates exergy efficiency, unit energy cost and carbon emission equivalent in real time. If the evaluation results deviate from the Pareto front by more than a predetermined distance for multiple consecutive evaluations, the model retraining process is triggered. The baseline curve for the daily plan is generated offline by cloud optimization, and the baseline curve is rolled over and corrected by the on-site student neural network model based on the measured data, forming a closed-loop optimization architecture. An abnormal operating condition identification mechanism is set in the closed-loop optimization architecture. An abnormal detection logic based on reconstruction error is adopted. If the reconstruction error of the input sensor vector after passing through the autoencoder exceeds the preset statistical upper limit, it will automatically switch to the conservative scheduling mode. In conservative dispatch mode, non-critical auxiliary loads are disconnected, the geothermal pump is locked at a safe reference speed to maintain basic heat energy supply, and the energy storage device is set to standby power mode.

Images