Water energy accumulator and ground source heat pump combined control method

Abstract

The invention discloses a water energy accumulator and ground source heat pump combined control method, relates to the technical field of building environment and equipment engineering, accurately captures a time sequence change rule of a building cold load based on the load prediction capability of a time convolution network, overcomes the defect that a traditional method is insufficient in response to sudden fluctuation, and improves the control efficiency. The system can plan an energy storage strategy in advance; an electricity price-load change rate sensitivity index is introduced, and through an adaptive weight adjustment mechanism, economic operation and load tracking requirements are dynamically balanced, so that strategy stiffness caused by fixed priorities is avoided, and the utilization efficiency of time-of-use electricity price signals is remarkably improved; according to the joint optimization design of the mixed integer linear programming model, the equipment operation constraint and the cost target are comprehensively considered, the collaborative decision of the energy accumulator charging and discharging strategy and the heat pump frequency setting is realized, and the direct energy supply dependence in the midday high electricity price period is effectively reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of building environment and equipment engineering technology, and particularly relates to a water accumulator combined with a ground source heat pump control method. BACKGROUND

[0002] Commercial buildings have significant cooling load peaks during the noon to afternoon period, and the time-of-use / peak-valley electricity price of the power grid makes the cost of electricity during the peak period higher. In order to reduce operating costs and smooth load fluctuations, the peak-shifting strategy of water storage and ground source heat pump (GSHP) cooperative operation is widely used: the GSHP charges the cold storage tank during the low-price period, and releases cold to the air conditioning system during the peak period, so as to reduce the unit power consumption and the maximum demand during the peak period. In recent years, load prediction and multi-objective optimization have been further introduced, the cooling load is predicted through a neural network model, and the charging / discharging time sequence is solved through a genetic algorithm. Some schemes use dynamic electricity price (time-of-use / day-ahead / real-time) signals to realize rolling adjustment, so as to take into account economy and energy efficiency.

[0003] However, the existing method still has the following problems in the scenario where load uncertainty and electricity price change coexist: firstly, prediction errors and load climbing rates often cause mismatch between the storage SOC and instantaneous demand, and the discharging power or coverage time is insufficient; secondly, some strategies rely on offline scheduling or fixed priorities, lack of rapid adaptation and feedback correction to sudden fluctuations, and are prone to cause frequent start-stop or high-frequency regulation of the GSHP; thirdly, the existing optimization does not explicitly depict key engineering boundaries such as the upper limit of water storage charging and discharging power, layered efficiency, GSHP buried pipe field heat balance, and maximum demand constraints. The above factors comprehensively limit the adjustment margin of the system under the coupling condition of the noon peak and electricity price constraints, and it is difficult to guarantee economy and comfort at the same time. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] The present application provides a water accumulator combined with a ground source heat pump control method to solve the problems of response lag, mismatch between energy storage and cold release capacity, and frequent start-stop of the heat pump caused by the dynamic coupling of the noon load peak and the electricity price of the power grid of commercial buildings.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] The present application provides a water accumulator combined with a ground source heat pump control method, which comprises:

[0008] Step S1, obtaining real-time data of building cooling load, indoor and outdoor temperature and humidity, and personnel occupancy proxy data, and a dynamic electricity price signal;

[0009] Step S2, performing short-term prediction on the cooling load based on a time convolution network to obtain a minute-by-minute prediction curve in a future control time domain;

[0010] Step S3, construct an electricity price-load rate of change sensitivity index, which weights and fuses the dynamic electricity price and the cooling load rate of change;

[0011] Step S4, establish a mixed integer linear programming model to jointly optimize the charging / discharging strategy of the water accumulator and the start / stop and frequency / number setting of the ground source heat pump, with the goal of minimizing the total operating cost;

[0012] Step S5, use a mixed integer linear programming solver to solve, to obtain the pump valve start / stop, compressor frequency or number, primary / secondary circulating water pump speed, etc. control instructions in the current control period;

[0013] Step S6, periodically roll over and update steps S2-S5, and trigger feedforward regulation to temporarily increase the cooling discharge capacity and correct the target state of charge when a load mutation event is detected.

[0014] As a preferred scheme of the water accumulator combined with the ground source heat pump control method, the input of the time convolution network includes historical cooling load sequence, indoor and outdoor temperature and humidity, and personnel occupancy agent data, and the time correlation feature extraction is enhanced through attention mechanism; the prediction time domain is 30-90 minutes, and a sliding window is used for online fine tuning.

[0015] As a preferred scheme of the water accumulator combined with the ground source heat pump control method, the electricity price-load rate of change sensitivity index is a linear combination of the two, and the weight coefficient is adaptively adjusted online through gradient descent with projection every control period, and the two dimensionless quantities are normalized and boundary constrained.

[0016] The online adaptive adjustment of the weight coefficient is as follows:

[0017] Step C1, give a unified online standardization and saturation processing for the dynamic electricity price and the load rate of change:

[0018] ,

[0019] ,

[0020] ,

[0021] ,

[0022] wherein, is the control period index, is the original value of the dynamic electricity price in the period, is the cooling load rate of change in the period, is the normalized dimensionless quantity, This is the moving average of the corresponding exponent. For the corresponding standard deviation estimate, As a smoothing factor, To trim the boundaries, This means that the same recursion is applied to both of these quantities. Indicates to Limited to Inside;

[0023] Step C2, construct weights based on normalized input and update them online:

[0024] ,

[0025] in, For the weight vector, The weights corresponding to electricity price and load change rate are respectively. For feasible sets, To The projection operator, Step size, This is the time smoothing coefficient. For the normalized input vector, This is the sensitivity scalar for this period. The weight of the previous period, For the updated weights, The normalized quantity used as a proxy;

[0026] Step C3, use positive part normalization to land the simplex projection while maintaining non-negativity and sum to one:

[0027] ,

[0028] in, Let be the vector to be projected. To take the non-negative values ​​by element, It is a vector of all ones.

[0029] As a preferred embodiment of the water storage device combined with ground source heat pump control method described in this invention, the total operating cost includes two parts: energy cost and maximum demand billing; the constraints include at least:

[0030] (1) The upper and lower limits of power, minimum start-up and shutdown time and power ramp-up rate of ground source heat pump;

[0031] (2) The upper and lower limits of the state of charge (SOC) of the water storage tank, the maximum charging / discharging power, and the prohibition of charging and discharging in the same cycle;

[0032] (3) Penalty for deviation between supply and return water temperatures and the comfortable range of indoor temperature;

[0033] (4) Optimize the consistency constraints between the time-domain rolling window and the control step size.

[0034] As a preferred embodiment of the water storage device combined with ground source heat pump control method described in this invention, the performance relationship of the ground source heat pump is characterized by a piecewise linearization method to adapt to mixed integer linear programming solution; and operating boundaries are set for ground source side temperature, well pump flow rate and reinjection temperature.

[0035] The piecewise linearization method for the performance relationship of the ground source heat pump is as follows:

[0036] Step D1: Construct a two-dimensional grid using compressor frequency and temperature difference; frequency nodes are spaced equidistantly within the rated range or densified according to historically occupied quantiles, while temperature difference nodes are densified in the low, medium, and high gradient regions of actual operation; the number of nodes depends on the solution time budget and accuracy requirements.

[0037] Step D2: Pre-label the performance table at the four vertices of each rectangular cell, and reconstruct the variables and performance energy using local convex combinations.

[0038] ,

[0039] in, Indicates the compressor frequency. This represents the temperature difference, specifically the temperature rise or fall between the supply water on the load side and the inlet water on the source side. Indicates the performance coefficient. Indicates cooling capacity. Indicates electrical power. and These represent the number of nodes in the direction of frequency and the direction of temperature difference, respectively. and For the cell index in both directions, These are the local indices of the four vertices of the cell. For vertex convex combination coefficients, For the binary variable of unit activation, , , , , These are the calibration constants for the corresponding vertices; from the same set Simultaneous reconstruction and This allows us to obtain a piecewise linear relationship diagram between the two for use in cost objectives;

[0040] Step D3: The linear equations and inequalities from step D2 are directly incorporated into the mixed integer linear programming. Only one cell can be activated, and the convex coefficients are summed to one within that cell. The operating boundary is added using linear inequalities: including the upper and lower limits of the source-side inlet and reinjection temperatures, the upper and lower limits of the well pump flow rate, and the allowable range of the secondary side supply and return water temperatures of the heat exchanger.

[0041] As a preferred embodiment of the water storage device combined with ground source heat pump control method described in this invention, the high / medium / low load ranges are identified online by a lightweight neural network classifier based on real-time load, predicted load and its rate of change, and a mode soft constraint or penalty coefficient is introduced into the optimization model accordingly: in the medium / low range, priority is given to cold storage or cold release, and in the high range, priority is given to direct supply and joint peak shaving is allowed.

[0042] As a preferred embodiment of the water storage combined with ground source heat pump control method described in this invention, the load change event is determined by the load change rate exceeding the threshold and / or a sudden increase in personnel occupancy; after triggering, the cooling upper limit of the water storage is temporarily increased and the SOC target bandwidth is adjusted, and the triggering duration and recovery strategy are set according to the event level classification.

[0043] As a preferred embodiment of the water storage device combined with ground source heat pump control method described in this invention, the solver adopts branch delimitation and uses the solution of the previous cycle as a hot start, and sets an upper limit of time and an optimal gap; if it is not feasible or there is no acceptable solution within the time limit, it switches to safety degradation control based on interval identification.

[0044] As a preferred embodiment of the water storage device combined with ground source heat pump control method described in this invention, the control commands include: the opening and closing sequence of relevant pump valves of the water storage device, the set value of the frequency or number of stages of the ground source heat pump compressor, the set value of the primary / secondary circulating water pump speed and the set value of the chilled water supply temperature, and include the constraint execution of the minimum start-up and shutdown time and speed ramp-up of the equipment.

[0045] As a preferred embodiment of the water storage combined with ground source heat pump control method described in this invention, the control cycle of rolling optimization is 5-15 minutes; and outlier elimination and missing measurement filling strategies are adopted for dynamic electricity price and load observation.

[0046] The beneficial effects of this invention are as follows: This invention constructs a collaborative optimization control framework for water storage devices and ground source heat pumps by integrating short-term load forecasting and dynamic electricity price sensitivity indicators.

[0047] This invention leverages the load forecasting capabilities of temporal convolutional networks to accurately capture the temporal variation patterns of building cooling load, overcoming the shortcomings of traditional methods in responding insufficiently to sudden fluctuations and enabling the system to plan energy storage strategies in advance. By introducing an electricity price-load change rate sensitivity index and employing an adaptive weight adjustment mechanism, it dynamically balances economic operation with load tracking requirements, avoiding strategy rigidity caused by fixed priorities and significantly improving the utilization efficiency of time-of-use electricity price signals. The joint optimization design of a mixed-integer linear programming model comprehensively considers equipment operating constraints and cost objectives, achieving coordinated decision-making between energy storage charging / discharging strategies and heat pump frequency settings, effectively reducing reliance on direct energy supply during peak midday electricity price periods. A periodic rolling update mechanism ensures that the control strategy is refreshed in real time according to changes in operating conditions, while event-triggered feedforward adjustment temporarily increases cooling capacity when sudden load changes are detected, enhancing the system's robustness to disturbances.

[0048] This invention optimizes the energy cost structure while maintaining indoor comfort, reduces losses caused by frequent equipment start-ups and shutdowns, and improves the system's economy and adaptability under dynamic electricity pricing. Simultaneously, the application of piecewise linearization and an online classifier balances solution efficiency and control accuracy, making complex optimization problems easier to deploy in engineering practice. This multi-level, adaptive control architecture provides a solution for commercial building energy systems that balances economy, stability, and response speed. Attached Figure Description

[0049] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0050] Figure 1 This is a schematic diagram of the control method for a water storage device combined with a ground source heat pump in the embodiment. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0053] For example, the terms “first” and “second” used in this application are only used to distinguish and describe similar objects, to differentiate the first object from another object, and are not used to describe a specific order or sequence, nor should they be interpreted as indicating or implying relative importance.

[0054] This application proposes a control method for a water storage device combined with a ground source heat pump, combining... Figure 1 As shown, the method includes:

[0055] Step S1: Obtain real-time building cooling load data, indoor and outdoor temperature and humidity data, occupancy proxy data, and dynamic electricity price signals (including at least one of time-of-use, day-ahead, and real-time).

[0056] Step S2: Short-term prediction of cooling load is performed based on a temporal convolutional network to obtain a minute-by-minute prediction curve in the future control time domain;

[0057] Step S3: Construct an electricity price-load change rate sensitivity index, which weights and integrates dynamic electricity price and cooling load change rate;

[0058] Step S4: With the goal of minimizing the total operating cost, establish a mixed integer linear programming model to jointly optimize the charging / discharging strategy of the water storage tank and the start / stop and frequency / stage settings of the ground source heat pump.

[0059] Step S5: Use a mixed integer linear programming solver to solve the problem and obtain control commands such as pump valve opening and closing, compressor frequency or number of stages, and primary / secondary circulating water pump speed in the current control cycle.

[0060] Step S6: Periodically and continuously update steps S2-S5. When a sudden load event is detected, trigger feedforward regulation to temporarily improve cooling capacity and correct the target state of charge, thereby achieving a balance between economy and responsiveness.

[0061] In this embodiment, the dynamic electricity price signal refers to any available source of time-of-use, day-ahead, or real-time electricity prices published according to the settlement cycle. The data is collected by the energy management system or metering endpoints and aligned with the control cycle. Personnel occupancy proxy data refers to observable measurements that reflect the strength of pedestrian flow, such as access control event counts or wireless access terminal counts, and is taken as the cumulative or average value within the control cycle. The default sampling cycle is 1 minute, and the control cycle is 10 minutes (configurable within 5-15 minutes), based on the metering system's publication frequency and the building automation system's refresh frequency. Data alignment adopts the caliber of the last snapshot or statistic within the previous complete control cycle. Optionally, if only time-of-use electricity prices are available but day-ahead / real-time electricity prices are not available, the time-of-use electricity price sequence is repeatedly expanded to the prediction time domain during optimization. If occupancy proxy data is temporarily unavailable, the occupancy intensity of the preset time period table or floor scheduling plan is used as a substitute input. If any input is missing in a single control cycle, the valid value of the previous cycle is used once. If the input is missing for two or more consecutive control cycles, a safety degradation is triggered, and the system operates only with available quantities and conservative boundaries until the data is recovered.

[0062] In one embodiment, the input to the temporal convolutional network includes historical cooling load sequences, indoor and outdoor temperature and humidity data, and occupancy proxy data, and the extraction of time-related features is enhanced through an attention mechanism; the prediction time domain is 30-90 minutes, and online fine-tuning is performed using a sliding window;

[0063] Specifically, the default time window for historical sequences is 60-120 minutes, and the window length is automatically set to 1-2 times this value depending on the prediction time domain. Online fine-tuning employs a small-step update strategy performed once per control cycle, with the number of training steps and learning rate dynamically converged based on the criterion that the prediction error does not increase for several consecutive cycles. The short-term prediction goal is to keep the mean absolute percentage error within 10-15% or significantly lower than the error level using the baseline from the same period yesterday. Optionally, if the online fine-tuning iteration fails to converge or resources are insufficient, it degenerates to fixed model inference and resumes online fine-tuning in the next cycle. If outliers exist in the input features, truncation or median replacement is performed before entering the model to avoid significant prediction shifts caused by single-cycle anomalies.

[0064] In one embodiment, the electricity price-load change rate sensitivity index is a linear combination of the two, and its weighting coefficient is adaptively adjusted online in each control cycle through gradient descent with projection, and the two dimensions are normalized and bounded.

[0065] The online adaptive adjustment method for the weighting coefficients is as follows:

[0066] Step C1 provides a unified online standardization and saturation process for dynamic electricity prices and load change rates:

[0067] ,

[0068] ,

[0069] ,

[0070] ,

[0071] in, To control the periodic index, This is the original value of the dynamic electricity price for this period. This represents the rate of change of cooling load during this cycle. For the normalized dimensionless quantity, This is the moving average of the corresponding exponent. For the corresponding standard deviation estimate, As a smoothing factor, To trim the boundaries, This means that the same recursion is applied to both of these quantities. Indicates to Limited to Inside;

[0072] Step C2, construct weights based on normalized input and update them online:

[0073] ,

[0074] in, For the weight vector, The weights corresponding to electricity price and load change rate are respectively. For feasible sets, To The projection operator, Step size, This is the time smoothing coefficient. For the normalized input vector, This is the sensitivity scalar for this period. The weight of the previous period, For the updated weights, The normalized quantity for the cost agent (its original quantity is composed of electricity price and load forecast increment, and is obtained by normalization in the same way as step C1).

[0075] Step C3, use positive part normalization to land the simplex projection while maintaining non-negativity and sum to one:

[0076] ,

[0077] in, Let be the vector to be projected. To take the non-negative values ​​by element, It is a vector of all ones;

[0078] Commonly used and feasible methods to take: , , , When observation noise is too high or weight fluctuations increase, improve Alternatively, using a decreasing step size can help suppress fluctuations; when the rolling period is 5-15 minutes, this value band matches the online update.

[0079] Specifically, the formulas listed here play a crucial role in unifying dimensions and adapting weights, and provide achievable simplex projections and numerical tuning bands for easy implementation. The normalization process uses recursion of the exponential moving average and variance, combined with symmetric pruning to limit feature amplitude and reduce the interference of outliers on the update direction. The weight update evolves within the simplex, and the physical meanings of non-negativity and summation remain stable. The time smoothing term suppresses excessive oscillations between adjacent periods. The projection implementation uses a closed-form expression with positive part normalization, which has low computational cost and is easy to embed within the controller cycle. The stability value range is guided by engineering adjustability, and recommended bands for step size, smoothing, and pruning are given. When field fluctuations are amplified, the trajectory can be further stabilized by increasing the smoothing coefficient or using a decreasing step size. For example, the normalized quantity of the cost proxy is obtained by combining the current electricity price and the predicted load increment after standardization with the same caliber. Its dimensions and amplitude are pruned to be comparable to the sensitivity scalar to ensure that the update direction is interpretable. The weights are initially evenly distributed and allowed to be biased after the first billing day. When the electricity price remains constant for several consecutive periods and the load change rate dominates, the weight evolution naturally converges to a combination dominated by the change rate. Optionally, when the weights are detected to oscillate repeatedly in a short period without improving the target value, the time smoothing term is temporarily increased and the weight control period is frozen for one cycle. If the weights cannot be stabilized for several consecutive periods, the weights are restored to the initial evenly distributed weights and the adaptive process is restarted. When any normalized input approaches zero due to an excessively small standard deviation, a minimum variance threshold replacement is used to avoid abnormally amplified update direction.

[0080] In one embodiment, the total operating cost includes both energy costs and maximum demand billing; the constraints include at least:

[0081] (1) The upper and lower limits of power, minimum start-up and shutdown time and power ramp-up rate of ground source heat pump;

[0082] (2) The upper and lower limits of the state of charge (SOC) of the water storage tank, the maximum charging / discharging power, and the prohibition of charging and discharging in the same cycle;

[0083] (3) Penalty for deviation between supply and return water temperatures and the comfortable range of indoor temperature;

[0084] (4) Optimize the consistency constraints between the time-domain rolling window and the control step size;

[0085] Similarly, the engineering constraints are as follows: the minimum start-up and shutdown time for ground source heat pumps is 10-15 minutes by default, and the power ramp-up rate is no more than 5-10% of the rated power per minute by default; the state-of-charge (SOC) operating range for hydroelectric storage is 10-90% by default, and the maximum charge / discharge power is given on the equipment nameplate or commissioning report. Simultaneous charge / discharge using mutually exclusive variables is prohibited. Comfort penalties are calibrated based on the allowable deviation of indoor temperature, with a default allowable deviation of ±0.5-1.0 degrees Celsius, and a penalty intensity that increases linearly with the deviation is set. Optionally, when maximum demand billing is not applicable, the corresponding terms can be masked in the objective function without affecting the feasibility of other constraints. If a device's state is close to the physical boundary, the corresponding decision variables will be constrained within the boundary during the current cycle, prioritizing the satisfaction of safety constraints.

[0086] In one embodiment, the performance relationship of the ground source heat pump is characterized by a piecewise linearization method to adapt to the solution of mixed integer linear programming; and operating boundaries are set for the ground source side temperature, well pump flow rate and reinjection temperature.

[0087] The piecewise linearization method for the performance relationship of ground source heat pumps is as follows:

[0088] Step D1: Construct a two-dimensional grid using compressor frequency and temperature difference (difference between load-side water supply and source-side inlet). Frequency nodes are spaced equidistantly within the rated range or densified according to historically occupied quantiles. Temperature difference nodes are densified in the low, medium, and high gradient regions of actual operation, prioritizing coverage of high load and high temperature difference areas. The number of nodes depends on the solution time budget and accuracy requirements, and 4-6 frequency nodes and 4-6 temperature difference nodes are commonly used to form a rectangular element.

[0089] Step D2: Pre-label the performance table at the four vertices of each rectangular cell, and reconstruct the variables and performance energy using local convex combinations.

[0090] ,

[0091] in, Indicates the compressor frequency. This represents the temperature difference, specifically the temperature rise or fall between the supply water on the load side and the inlet water on the source side. Indicates the performance coefficient. Indicates cooling capacity. Indicates electrical power. and These represent the number of nodes in the direction of frequency and the direction of temperature difference, respectively. and For the cell index in both directions, These are the local indices of the four vertices of the cell. For vertex convex combination coefficients, For the binary variable of unit activation, , , , , These are the calibration constants for the corresponding vertices; from the same set Simultaneous reconstruction and This allows us to obtain a piecewise linear relationship diagram between the two for use in cost objectives;

[0092] Step D3: The linear equations and inequalities from step D2 are directly incorporated into the mixed integer linear programming. Only one cell can be activated, and the convex coefficients are summed to one within that cell. The operating boundary is added with linear inequalities: including the upper and lower limits of the source-side inlet and reinjection temperatures, the upper and lower limits of the well pump flow rate, and the allowable range of the secondary side supply and return water temperatures of the heat exchanger. If necessary, soft penalties are set in the temperature difference direction to cover feasible operations outside the extreme points.

[0093] The vertex constant is derived from the manufacturer's performance curve or simulation calibration. When seasonal drift is detected, the table entries can be recalibrated offline at fixed intervals and replaced. The node density remains unchanged to stabilize the solution size.

[0094] Specifically, a two-dimensional piecewise linearization framework with frequency and temperature difference as independent variables is used here. Performance mapping is achieved within MILP through element activation and vertex convex combination. This variable pair is chosen to centrally reflect the thermodynamic improvement and control accessibility of the ground source heat pump, reducing dimensionality while retaining key gradients. Performance entries are pre-stored at the four vertices of each element, and frequency, temperature difference, coefficient of performance, cooling capacity, and electrical power are reconstructed simultaneously with unified convex coefficients. This transforms the objective function and constraints into linear expressions, avoiding nonlinear coupling caused by reciprocals or products, thus maintaining the solution difficulty at an acceptable level. Node selection follows high gradient densification and commonly used areas. The principle of inter-rate density is adopted, balancing computational efficiency and approximation quality. The operating boundary is uniformly incorporated using linear inequalities, maintaining consistency with ground source temperature, well pump flow rate, and reinjection limits. This structure is maintainable against field drift, requiring only vertex table updates without altering the model's morphology. Optionally, vertex table calibration prioritizes manufacturer performance data, with sampling verification at field steady-state measurement points. The verification calibrator ensures that the relative errors of cooling capacity and electrical power are no higher than 5-7% at sampling points covering common operating conditions. When observations show persistently large errors during a certain seasonal period, offline recalibration is performed monthly or quarterly, replacing table entries without changing the grid density. Two-dimensional grids are typically 5×5 or approximate. When operating points fall outside grid coverage, they are first projected onto the nearest boundary node at the variable level, with soft penalties ensuring feasibility and convergence. If any vertex entry is missing or deemed abnormal, interpolated values ​​from adjacent vertices are temporarily used, and the missing entry is corrected during the next calibration.

[0095] In one embodiment, a lightweight neural network classifier identifies high / medium / low load zones online based on real-time load, predicted load, and their rate of change. This identification introduces pattern soft constraints or penalty coefficients into the optimization model: prioritizing cooling storage or release in medium / low load zones, and prioritizing direct supply and allowing joint peak shaving in high load zones, with the final output determined by a unified optimization decision. Furthermore, the zone identification classifier employs a small-scale network structure to meet real-time requirements within the control cycle. The input is a concatenated vector of load, prediction, and rate of change from the current and several neighboring cycles. By default, each classification result must remain consistent across two consecutive control cycles to be considered a valid basis for pattern soft constraints, thus avoiding pattern jitter. Online updates to the classifier are performed in batches daily or weekly by default, with updates triggered only if the validation set accuracy does not decrease. Optionally, when the classifier output confidence is too low or conflicting, it degenerates to a penalty setting for the medium load mode and remains until the next cycle for re-evaluation. If the classifier component is unavailable, the optimization solves directly based on the objective and physical constraints without using pattern soft constraints.

[0096] In one embodiment, a load surge event is determined by the load change rate exceeding the threshold and / or a sudden increase in personnel occupancy. Upon triggering, the upper limit of cooling capacity of the water storage tank is temporarily increased, and the SOC target bandwidth is adjusted. The trigger duration and recovery strategy are set according to the event level. In this embodiment, the load surge rate threshold is set based on the installed cooling capacity and historical fluctuation characteristics. By default, a valid event is determined when the relative increase within a single control cycle reaches a certain percentage point and lasts for at least one control cycle. A sudden increase in personnel occupancy is based on the increase in agent volume per unit time and is weighted and summarized by floor or region. After the event is triggered, the increase in the cooling capacity limit does not exceed the transient power limit allowed by the equipment nameplate, and the SOC target bandwidth is moderately reduced without touching the lower limit safety line. Optionally, the event cancellation criterion is that the load change rate falls back below the threshold and remains below it for several consecutive control cycles. After cancellation, the normal target bandwidth is restored linearly to avoid secondary disturbances. If the number of consecutive triggers is too high, a suppression window is entered, within which the trigger threshold is increased to avoid excessively frequent feedforward actions.

[0097] In one embodiment, the solver employs branch-bound and uses the solution from the previous cycle as a warm start, sets a time limit and an optimal gap; if it is infeasible or there is no acceptable solution within the time limit, it switches to interval-based safety degradation control.

[0098] For example, safety degradation control prioritizes using feasible solutions from the previous cycle and limits changes in power and speed within the current cycle to no more than the ramp-up limit, while maintaining indoor temperature within the comfort zone boundary. If the solution from the previous cycle is unavailable, a fixed output mode based on interval identification is adopted, and cold storage is prioritized to ensure short-term feasibility while the hydroelectric accumulator's SOC remains within the safe range. The time limit for solving mixed-integer linear programming is set to 30-60 seconds by default, and the optimal gap is set to 2-5% by default, based on the control cycle and computational resources. Hot start uses the variable values ​​from the previous cycle as the initial point. If an acceptable solution is not obtained within the time limit, the aforementioned degradation strategy is immediately executed and an alarm is recorded; the normal solution process resumes in the next cycle.

[0099] In one embodiment, the control commands include: the opening and closing sequence of pumps and valves related to the water storage tank, the set value of the frequency or number of stages of the ground source heat pump compressor, the set value of the speed of the primary / secondary circulating water pump and the set value of the chilled water supply temperature, and include the execution of constraints on the minimum start-up and shutdown time and speed ramp-up of the equipment.

[0100] Similarly, before issuing instructions, the execution layer performs legality and interlock checks on the status of critical equipment to ensure that pump and valve instructions do not conflict with the minimum start-up and shutdown times of the equipment. Instruction issuance adopts a standard of sending immediately upon completion of calculation and aligning the execution effective time with the boundary of the next control cycle. The default delay for execution confirmation is no more than a certain number of seconds; if this time limit is exceeded, it is considered an execution failure and a retry is performed. Optionally, the adjustment of the chilled water supply temperature setpoint adopts a discrete step method, with each step not exceeding a certain number of divisions to avoid secondary side fluctuations. If equipment rejection or failure to confirm within the time limit is received, the current safe state is maintained and reassessed in the next cycle.

[0101] In one embodiment, the control cycle for rolling optimization is 5-15 minutes. Outlier removal and missing data filling strategies are employed for dynamic electricity prices and load observations to ensure the stability of online adaptive and feedforward criteria. Optionally, outlier removal uses a robust statistical method based on control cycles, with a default window length of 3-5 control cycles. The replacement strategy prioritizes using the median of adjacent cycles or the previous valid value. Missing data filling uses the most recent observation for a single missing data instance; if consecutive missing data exceeds two control cycles, a safety degradation is triggered, and online adaptive updates are stopped until data recovery. Clock synchronization and timestamp alignment use the energy management system as the master clock, allowing drift of no more than a few seconds. When the release time of electricity price or occupancy data lags behind the control cycle, the data from the latest complete cycle is used in the current solution and automatically compensated in the next cycle. When dynamic electricity prices switch across days or billing windows switch, the settings before the switch are used within that cycle to avoid double billing illusions, and calculations are performed according to the new window in the next cycle.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0103] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A water storage device combined with a ground source heat pump control method for use in commercial building air conditioning, characterized in that, include: Step S1: Obtain real-time data on building cooling load, indoor and outdoor temperature and humidity, occupancy data, and dynamic electricity price signals; Step S2: Short-term prediction of cooling load is performed based on a temporal convolutional network to obtain a minute-by-minute prediction curve in the future control time domain; Step S3: Construct an electricity price-load change rate sensitivity index, which weights and integrates dynamic electricity price and cooling load change rate; Step S4: With the goal of minimizing the total operating cost, establish a mixed integer linear programming model to jointly optimize the charging / discharging strategy of the water storage tank and the start / stop and frequency / stage settings of the ground source heat pump. Step S5: Use a mixed integer linear programming solver to solve the problem and obtain control commands such as pump valve opening and closing, compressor frequency or number of stages, and primary / secondary circulating water pump speed in the current control cycle. Step S6: Periodically update steps S2-S5. When a load change event is detected, trigger feedforward adjustment to temporarily improve cooling capacity and correct the target state of charge.

2. The water storage device combined with a ground source heat pump control method as described in claim 1, characterized in that, The input to the temporal convolutional network includes historical cooling load sequences, indoor and outdoor temperature and humidity data, and occupancy proxy data. It enhances the extraction of time-related features through an attention mechanism. The prediction time domain is 30-90 minutes, and online fine-tuning is performed using a sliding window.

3. The water storage device combined with a ground source heat pump control method as described in claim 1, characterized in that, The electricity price-load change rate sensitivity index is a linear combination of the two, and its weighting coefficient is adaptively adjusted online in each control cycle through gradient descent with projection, and the two dimensions are normalized and bounded. The online adaptive adjustment method for the weight coefficients is as follows: Step C1 provides a unified online standardization and saturation process for dynamic electricity prices and load change rates: ， ， ， ， in, To control the periodic index, This is the original value of the dynamic electricity price for this period. This represents the rate of change of cooling load during this cycle. For the normalized dimensionless quantity, This is the corresponding exponential moving average. For the corresponding standard deviation estimate, As a smoothing factor, To trim the boundaries, This indicates that the same recursion is applied to both of these quantities. Indicates to Limited to Inside; Step C2, construct weights based on normalized input and update them online: ， in, For the weight vector, The weights corresponding to electricity price and load change rate are respectively. For feasible sets, To The projection operator, Step size, This is the time smoothing coefficient. For the normalized input vector, This is the sensitivity scalar for this period. The weight of the previous period, For the updated weights, The normalized quantity used as a proxy; Step C3, use positive part normalization to land the simplex projection while maintaining non-negativity and sum to one: ， in, Let be the vector to be projected. To take non-negative values ​​by element, It is a vector of all ones.

4. The water storage device combined with a ground source heat pump control method as described in claim 1, characterized in that, The total operating cost includes two parts: electricity cost and maximum demand billing; the constraints include at least: (1) The upper and lower limits of power, minimum start-up and shutdown time and power ramp-up rate of ground source heat pump; (2) The upper and lower limits of the state of charge (SOC) of the water storage tank, the maximum charging / discharging power, and the prohibition of charging and discharging in the same cycle; (3) Penalty for deviation between supply and return water temperatures and the comfortable range of indoor temperature; (4) Optimize the consistency constraints between the time-domain rolling window and the control step size.

5. The water storage device combined with a ground source heat pump control method as described in claim 1, characterized in that, The performance relationship of the ground source heat pump is characterized by a piecewise linearization method to adapt to the solution of mixed integer linear programming; and operating boundaries are set for the ground source temperature, well pump flow rate and reinjection temperature. The piecewise linearization method for the performance relationship of the ground source heat pump is as follows: Step D1: Construct a two-dimensional grid using compressor frequency and temperature difference; frequency nodes are spaced equidistantly within the rated range or densified according to historically occupied quantiles, while temperature difference nodes are densified in the low, medium, and high gradient regions of actual operation; the number of nodes depends on the solution time budget and accuracy requirements. Step D2: Pre-label the performance table at the four vertices of each rectangular cell, and reconstruct the variables and performance energy using local convex combinations. ， in, Indicates the compressor frequency. This represents the temperature difference, specifically the temperature rise or fall between the supply water on the load side and the inlet water on the source side. Indicates the performance coefficient. Indicates cooling capacity. Indicates electrical power. and These represent the number of nodes in the direction of frequency and the direction of temperature difference, respectively. and For the cell index in both directions, These are the local indices of the four vertices of the cell. For vertex convex combination coefficients, For the binary variable of unit activation, , , , , These are the calibration constants for the corresponding vertices; from the same set Simultaneous reconstruction and This allows us to obtain a piecewise linear relationship diagram between the two for use in cost objectives; Step D3: The linear equations and inequalities from step D2 are directly incorporated into the mixed integer linear programming. Only one cell can be activated, and the convex coefficients are summed to one within that cell. The operating boundary is added using linear inequalities: including the upper and lower limits of the source-side inlet and reinjection temperatures, the upper and lower limits of the well pump flow rate, and the allowable range of the secondary side supply and return water temperatures of the heat exchanger.

6. The water storage device combined with a ground source heat pump control method as described in claim 1, characterized in that, Based on real-time load, predicted load and its rate of change, a lightweight neural network classifier identifies high / medium / low load ranges online, and accordingly introduces mode soft constraints or penalty coefficients into the optimization model: prioritizing cold storage or release in medium / low ranges, and prioritizing direct supply and allowing joint peak shaving in high ranges.

7. The water storage device combined with a ground source heat pump control method as described in claim 1, characterized in that, Load mutation events are determined by load change rate exceeding the threshold and / or sudden increase in personnel occupancy; after triggering, the upper limit of cooling of the water storage tank is temporarily increased and the SOC target bandwidth is adjusted, and the trigger duration and recovery strategy are set according to the event level classification.

8. The water storage device combined with a ground source heat pump control method as described in claim 1, characterized in that, The solver employs branch-and-bound and uses the solution from the previous cycle as a warm start, setting a time limit and an optimal gap; if it is not feasible or there is no acceptable solution within the time limit, it switches to safety degradation control based on interval identification.

9. A water storage device combined with a ground source heat pump control method as described in claim 1, characterized in that, The control commands include: the opening and closing sequence of pumps and valves related to the water storage tank, the set value of the frequency or number of stages of the ground source heat pump compressor, the set value of the speed of the primary / secondary circulating water pump and the set value of the chilled water supply temperature, and include the constraint execution of the minimum start-up and shutdown time and speed ramp-up of the equipment.

10. A water storage device combined with a ground source heat pump control method as described in claim 1, characterized in that, The control cycle for rolling optimization is 5-15 minutes; outlier removal and missing data filling strategies are adopted for dynamic electricity prices and load observations.

Images