High-voltage DC power supply method and system based on LSTM prediction

Abstract

The invention belongs to the technical field of power supply, and relates to a high-voltage direct-current power supply method and system based on LSTM prediction, and the method comprises the following steps: obtaining historical and current-day electricity price data, constructing an LSTM model, and predicting the direct power supply cost of a next time slot; establishing a joint objective function, and calculating the power supply cost of each storage battery pack; determining a power supply source and power distribution of the current time slot based on the direct power supply cost and the power supply cost of each storage battery pack; controlling alternating current and direct current isolation on-off of the corresponding sections, establishing or disconnecting an energy channel between each storage battery pack and a bus, and generating a charging and discharging plan; and updating the direct power supply cost and the charging and discharging plan based on the new observation data before the next time slot is reached. According to the technical scheme, the LSTM prediction model is introduced into the high-voltage direct-current power supply system to dynamically predict the electricity price in the target time period, cost quantification and look-ahead decision making for different power supply paths are achieved, and a power supply control strategy is made to have self-adaptability and time sequence look-ahead.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power supply, and particularly relates to a high-voltage direct-current power supply method and system based on LSTM prediction. BACKGROUND

[0002] In the prior art, the power supply of a data center usually depends on an UPS architecture (Uninterruptible Power Supply Architecture), and power stability is maintained through a link of commercial power-UPS-server. However, the UPS architecture is difficult to meet the comprehensive requirements of high power density, continuous power supply and economic operation of an AI data center because of the decline of the chemical activity of the battery due to long-term floating charge, long discharge maintenance period, low energy recovery efficiency and generally insufficient power utilization.

[0003] With the emergence of an HVDC architecture (High Voltage Direct Current Architecture), some systems attempt to improve efficiency by reducing the AC / DC (Alternating Current / Direct Current) conversion link, but in most existing solutions, the HVDC system usually adopts a fixed topology and a static charging and discharging strategy, and has limited response capability to price fluctuations, and it is difficult to dynamically adjust the power supply role of the energy storage unit, so that the energy storage device has a delay and redundant power supply problem when charging and discharging is switched.

[0004] At the same time, traditional price prediction or energy storage scheduling usually relies on artificial experience or static models, and it is difficult to capture the long-term and short-term change rules of the price sequence, so that the scheduling result lags behind the market price signal, and the energy storage device cannot perform charging and discharging operation in the optimal price window, and the system economy and energy efficiency are restricted.

[0005] Therefore, the prior art usually has problems such as power supply cost response lag, energy storage scheduling rigidity and insufficient price prediction accuracy. This is the deficiency of the prior art.

[0006] Therefore, it is necessary to provide a high-voltage direct-current power supply method and system based on LSTM prediction to solve the above-mentioned defects in the prior art. SUMMARY

[0007] The purpose of the present application is to provide a high-voltage direct-current power supply method and system based on LSTM prediction to solve the above-mentioned technical problems.

[0008] To achieve the above object, the present application provides the following technical solutions: A high-voltage direct current power supply method based on LSTM prediction, comprising the following steps: Obtain historical and current day electricity price data, build an LSTM model, use the LSTM model to predict the electricity price data of the target period, and obtain the direct power supply cost of the next time slot; Establish a joint objective function, calculate the power supply cost of each battery pack, and the parameters of the joint objective function include at least one of the charging electricity price, the battery charging and discharging efficiency, the bus power constraint and the energy storage life constraint; Based on the direct power supply cost and the power supply cost of each battery pack, determine the power supply source and power distribution of the current time slot in the topology of the segmented high-voltage direct current bus; Control the on-off of the AC rectifier and the on-off of the DC isolation of the corresponding segment, establish or break the energy channel between each battery pack and the bus, and generate the charging and discharging plan; Before reaching the next time slot, update the prediction of the direct power supply cost and the charging and discharging plan based on new observation data.

[0009] The above technical solutions are adopted, the LSTM (Long Short-Term Memory, long short-term memory network) prediction model is introduced into the high-voltage direct current power supply system to dynamically predict the electricity price of the target period, the cost quantification and forward decision of different power supply paths are realized, the power supply control strategy has self-adaptability and time sequence foresight, and the demand for timely response of data center power supply cost, flexibility of energy storage scheduling and reliable prediction accuracy is met.

[0010] The LSTM model is built by obtaining historical and current day electricity price data, so that the electricity price prediction is no longer dependent on fixed assumptions and has long-term and short-term trend separation capability, providing continuity support in time dimension for power supply cost determination; by establishing a joint objective function, the charging electricity price, the battery charging and discharging efficiency and the energy storage life constraint are unified into the optimization target, so that the calculation of the power supply cost not only reflects the economic factors, but also considers the safety of energy storage recycling and power boundary; by comparing the direct power supply cost and the battery power supply cost in the topology of the segmented high-voltage direct current bus, the power supply source and power distribution are determined, so that the system energy supply path can balance between economic optimization and capacity constraint; by controlling the on-off of the AC rectifier and the on-off of the DC isolation of each segment, the energy channel between the battery pack and the bus is established or broken, realizing real-time switching and coordinated operation between different energy units; by rolling updating the prediction model and the charging and discharging plan according to new observation data before the next time slot, the control instruction can continuously maintain dynamic consistency with market signals and operating state, so as to realize the unification of economic scheduling and high-reliability power supply, and meet the demand for low-cost, continuity and intelligent power supply of high-power density data centers.

[0011] As a preferred approach, the inputs to the LSTM model include at least one of the following: historical electricity price data, current electricity price data, distributed power output, real-time data center load, and battery state of charge. The LSTM model introduces a gating mechanism in the hidden layer output to model long-term trends and short-term fluctuations separately.

[0012] This technical solution introduces multi-source operating parameters into the input of the LSTM model, enabling the prediction model to establish a dynamic mapping between electricity prices and system operating status within a multi-dimensional data feature space, achieving the following technical effects: First, while receiving historical electricity price data and current electricity price data, the model takes information such as distributed power output, real-time load and battery charge status as variable inputs, so that the prediction process takes into account both time series changes and energy supply and demand characteristics, forming an overall response capability to electricity price trends, thereby maintaining prediction stability in the scenario of electricity price fluctuations. Second, the gating mechanism separates the long-term trend and short-term fluctuations in the hidden layer output, enabling the model to maintain trend memory in the long-term series and have the ability to adjust quickly under short-term disturbances. This avoids the prediction bias and cumulative error caused by single time scale modeling and improves the adaptability of the prediction results to time series changes. Third, the model, through the synergistic effect of multiple inputs and gating mechanisms, achieves the identification of time-series dependencies between power supply-related variables in the data center. This enables the predicted output to more accurately reflect the direction and fluctuation range of electricity price changes within future time slots, providing a high-precision reference for subsequent power supply cost calculation and scheduling decisions, and enhancing the system's response reliability and economy in dynamic market environments.

[0013] Preferably, the joint objective function is also used to calculate the available power of each battery pack, and to keep the sum of the available power of each battery pack from being lower than a preset critical value under the constraint of available power.

[0014] This technical solution incorporates available power constraints into the joint objective function, combining the battery pack's energy state with the power supply cost calculation process. This enables the optimization decision-making to possess capacity self-balancing capabilities and energy storage redundancy management characteristics, achieving the following technical effects: First, the joint objective function outputs the available power of each battery pack while calculating the power supply cost, so that the cost optimization result not only depends on the electricity price prediction, but also is tied to the actual energy storage capacity, thereby avoiding the deep discharge and capacity imbalance problems caused by simply relying on economic judgment, and improving the long-term operational stability of the energy storage unit. Second, by maintaining the sum of available power of each battery pack at a preset threshold under the constraint of available power, the system always has dispatchable energy storage redundancy during operation. When the main power supply fluctuates or the electricity price rises abnormally, it can still maintain the normal power supply path, thus achieving the parallel development of power supply continuity and energy security. Thirdly, the constraint condition is embedded as a hard constraint in the joint objective function solution, so that the optimization process realizes dynamic balance between global cost minimization and energy storage safety boundary, avoids life attenuation caused by frequent charging and discharging of a single battery group, and improves the stability of power scheduling and the balanced utilization rate of the energy storage group, thereby maintaining long-term consistency of power supply economy and reliability.

[0015] Preferably, the determination of the power supply source is based on the comparison result of the direct power supply cost and the power supply cost of each battery group, and the power supply source is selected according to the cost minimization principle in the current time slot.

[0016] In this technical solution, the cost minimization principle is introduced in the power supply decision, and a dynamic decision mechanism is established based on the comparison result of the direct power supply cost and the power supply cost of each battery group, so that the power supply path selection has self-optimization characteristics and can achieve the following technical effects: Firstly, the selection of the power supply source takes the cost parameter of the current time slot as the judgment core, quantitatively compares the economy of transformer rectification power supply and multi-battery group power supply, makes the power supply path match the lowest cost scheme in real time, reduces unnecessary energy conversion loss, and improves the overall power supply efficiency. Secondly, under the action of the cost minimization principle, the charging and discharging behavior of different battery groups and the rectification power supply strategy form a complementary relationship, so that the system can automatically switch the power supply mode under the scenes of electricity price fluctuation and load change, avoid the redundant operation and power waste in traditional static scheduling, and improve the flexibility of system energy distribution. Thirdly, the dynamic decision of the power supply source enables the high-voltage DC bus in the segmented topology to realize rapid adjustment according to the equivalent cost of each path, taking into account power balance and economic optimization, so that the energy maintains the optimal operating state when flowing between multiple paths, thereby realizing real-time economic scheduling and high-reliability operation of the power supply system, and improving the overall energy utilization rate and the intelligent level of scheduling.

[0017] Preferably, the determination of the power distribution is based on the difference between the direct power supply cost and the power supply cost of each battery group to establish a distribution weight, and the power output is distributed according to the topological relationship of the segmented high-voltage DC bus.

[0018] In this technical solution, the weight construction method based on the cost difference is introduced in the power distribution link, and the power output is distributed in combination with the topological relationship of the segmented high-voltage DC bus, so that the power distribution process changes from static control to dynamic economic scheduling, and the following technical effects can be achieved: Firstly, the cost difference is used to establish a distribution weight, so that the power distribution ratio is directly related to the economy of each power supply path, thereby realizing optimal distribution of energy flow while keeping the total power constant under different power supply modes, reducing the energy proportion of high-cost paths, and improving the overall energy efficiency of the system. Second, by combining the topology relationship of the segmented high-voltage DC bus to distribute power output, the segmented bus can maintain load balance in different operating states, avoid voltage fluctuation and energy loss caused by centralized power supply or single bus overload, and realize collaborative power supply between segments; Third, the dynamic update of the power distribution weight enables the system to adjust the energy output structure according to real-time cost changes, making the bus power flow have time sequence adjustability and spatial coordination, thereby ensuring balanced discharge of energy storage units and efficient use of rectification paths, and enabling the power supply network to maintain stable, efficient and economic operation in a complex electricity price environment.

[0019] As a preferred, the control of the AC rectification on-off and DC isolation on-off of the corresponding segment includes: determining the switching time according to the energy remaining rate and the electricity price change rate of each segment, and using a delay confirmation strategy to determine the switching time, and keeping the on-off state unchanged when the difference between the adjacent time slot electricity price prediction is lower than the set threshold.

[0020] In this technical solution, the energy remaining rate and the electricity price change rate are combined as the on-off criterion, and a delay confirmation strategy is set to suppress short-term fluctuations, making the on-off control have threshold memory and switching lag characteristics, which can achieve the following technical effects: First, on-off switching is only triggered when both the energy state and the price signal meet the criterion, avoiding frequent transitions caused by a single factor, maintaining the stable on-off relationship of the power channel, and reducing the conversion loss and thermal stress caused by the high-frequency operation of the rectifier and isolator; Second, the delay confirmation maintains the original on-off state when the difference between the adjacent time slot prediction is small, so that small prediction noise no longer triggers topology shock, reduces the repeated rise and fall of bus voltage and current, and maintains the continuity of segmented load distribution and the coherence of scheduling plan; Third, the energy remaining rate participates in the criterion, which makes different segments prefer to maintain stable channels near the capacity boundary, reduces the shallow cycle life consumption of energy storage units caused by short-term switching, and makes the on-off rhythm consistent with the rolling plan, improving the predictability and economy of the entire power supply control in multi-time slot operation.

[0021] As a preferred, the generation of the charge and discharge plan includes setting a rotation priority sequence for each battery pack, and the rotation priority sequence is determined based on the cycle number, current temperature rise and state of charge of the battery pack.

[0022] In this technical solution, the rotation priority sequence is constructed based on the cycle number, temperature rise and state of charge, and the power-on and standby order of each battery pack is arranged accordingly, so that the scheduling changes from static fixed to adaptive rotation based on state variables, which can achieve the following technical effects: First, the rotation priority sequence allows high-frequency units to give way to low-frequency units in time, the depth of discharge and cumulative cycles are distributed among units, excessive concentration of use is inhibited, capacity degradation is more balanced, standby capacity remains in a stable range, and energy redundancy can be called upon in abnormal fluctuations, power continuity and scheduling predictability are strengthened; Second, the temperature rise is used as a key input to make the thermal constraint explicit, the thermal load is actively dispersed in rotation, the thermal gradient and hot spot formation of each section of the bus are suppressed, the thermal stress accumulation rate of insulation and conductor is reduced, the internal resistance change during charging and discharging is smoother, the power command response is more stable, the voltage drop and rise amplitude is reduced, and the thermal-related protection action triggering frequency is reduced. Third, combined with the state of charge, the priority is evaluated, different units around the appropriate SOC (State of Charge) interval alternately undertake power output and charging tasks, the deep and shallow cycle ratio is more reasonable, the equivalent loss during energy path switching is reduced, the power distribution and capacity boundary are kept in coordination, and the plan rolling adjustment under price and load disturbance is more smooth, the comprehensive operation cost and available capacity utilization rate show stable income.

[0023] As a preferred, the update is performed by incremental learning of LSTM model parameters in each prediction period, the model weight is continuously updated using a sliding window mechanism, and the prediction period is adaptively shortened when the electricity price fluctuates sharply.

[0024] In the technical scheme, the incremental learning of model parameters is performed in the prediction period, and the sliding window update mechanism is combined, so that the LSTM model can continuously maintain the sensitivity and generalization ability to the time series characteristics of the electricity price, adaptively adjust the prediction period in the sharp fluctuation scenario, and the following technical effects can be achieved: First, the incremental learning mechanism updates only the local weight when receiving new observation data, retains the stability of the original structure, maintains the convergence characteristics in the long-term operation, and gradually absorbs the latest price change information, so that the prediction accuracy is continuously optimized over time, and the performance degradation of the model caused by static parameters is avoided. Second, the sliding window mechanism ensures that the training samples are continuously rolling in time sequence, the weights of new and old data are dynamically balanced, the overfitting of recent abnormalities is prevented, and the trend migration process is captured, so that the prediction output forms a stable response between short-term fluctuations and long-term trends, and the smoothness and reliability of the prediction curve are maintained. Third, the prediction period is adaptively shortened when the electricity price fluctuates sharply, so that the model update frequency matches the market change rhythm, reduces the decision deviation caused by prediction lag, and makes the power supply scheduling plan still have timeliness and economy in the high-speed fluctuation interval, realizes the dynamic adaptation ability and continuous efficient operation of the prediction model in the complex market environment.

[0025] Furthermore, the application also provides a high-voltage direct-current power supply system based on LSTM prediction, comprising: a power price prediction module, configured to construct an LSTM model based on historical power price data and daily power price data, and to predict power price data of a target period to obtain direct power supply cost of a next time slot; a cost calculation module, configured to establish a joint objective function to calculate power supply cost of each battery pack, parameters of the joint objective function including at least one of charging price, battery charging and discharging efficiency, bus power constraint and energy storage life constraint; a power supply decision module, configured to determine power supply source and power distribution of the current time slot in the topology of the segmented high-voltage direct-current bus according to the direct power supply cost and the power supply cost of each battery pack; a topology control module, configured to control on-off of alternating current rectification and direct current isolation of the segmented high-voltage direct-current bus according to an output signal of the power supply decision module, and to establish or disconnect an energy channel between each battery pack and the bus; a plan generation module, configured to generate and store a charging and discharging plan of the battery pack, the charging and discharging plan being updated synchronously with a switching instruction of the power supply source; a parameter update module, configured to update the power price prediction module and the charging and discharging plan based on new observation data before reaching the next time slot.

[0026] By adopting the above technical solution, the power price prediction module, the cost calculation module, the power supply decision module, the topology control module, the plan generation module and the parameter update module are arranged in the high-voltage direct-current power supply system, so that prediction driving and multi-source collaboration of power supply control are realized, and the power supply path is optimized and the system operation is efficient under dynamic power price and complex load conditions.

[0027] The power price prediction module constructs an LSTM model by using historical and daily power price data, so that the system has time sequence learning and feedforward prediction ability for power price trend, and provides real-time basis for subsequent power supply cost determination; the cost calculation module models charging price, battery efficiency, power constraint and life parameters by using a joint objective function, so as to realize quantitative evaluation and economic judgment of power supply cost; the power supply decision module selects the optimal power supply path and power distribution based on the cost comparison result, so that the energy dispatching process is converted from fixed logic to real-time optimization; the topology control module executes on-off instructions of alternating current rectification and direct current isolation channels according to the decision signal, so that the energy flow direction of each battery pack is automatically reconstructed according to the dispatching result, and the power transmission path and operation strategy are consistent; the plan generation module records the charging and discharging sequence and keeps it synchronous with the power supply switching instruction, so that the energy storage behavior is coordinated with the market price rhythm; and the parameter update module performs rolling correction on model parameters and dispatching plans before each time slot, so that the system continuously maintains prediction accuracy and dispatching consistency in long-term operation, thereby realizing comprehensive optimization of cost, efficiency and stability of the high-voltage direct-current power supply of the data center.

[0028] As preferred, the power supply decision module comprises a power distribution unit for calculating a distribution weight according to the difference between the direct power supply cost and the battery pack power supply cost, and generating a power distribution instruction according to the distribution weight.

[0029] In the power supply decision module, the weight calculation mechanism based on the cost difference is introduced in the power distribution unit, so that the generation of the power instruction takes into account the coordination of economy and electrical constraints, and the following technical effects can be achieved: First, the power distribution unit establishes a weight matrix for multiple energy supply paths with the cost difference as the decision quantity, so that the output power of each path is directly linked to its economic contribution, thereby automatically reducing the high-cost channel and automatically increasing the low-cost channel, forming a dynamic self-balancing power supply distribution process that changes with market signals, reducing redundant energy flow and unnecessary power repeated switching; Second, the weight generation mechanism enables the power distribution result to be synchronized with the bus topology structure, and the power output ratio between different battery packs is adjusted in real time according to their respective capacities, efficiencies and bus constraints, avoiding uneven energy efficiency caused by overloading or idling of a single battery, and improving the overall power utilization rate and bus operation stability of the system; Third, the power distribution instruction is issued by the power supply decision module after real-time calculation, so that the system can quickly respond and reconstruct the power distribution pattern under conditions of load mutation or sharp price fluctuations, maintain continuous output and voltage balance, and improve the adaptability and sustained economic operation capability of the high-voltage direct current power supply system under complex operating conditions.

[0030] The beneficial effects of the present application are that by introducing an LSTM prediction model in the high-voltage direct current power supply system to dynamically predict the target period price, the cost of different power supply paths is quantified and forward decision is realized, so that the power supply control strategy has adaptability and time sequence foresight, and meets the needs of timely response of data center power supply cost, flexible energy storage scheduling and reliable prediction accuracy.

[0031] In addition, the design principle of the present application is reliable, the structure is simple, and it has very wide application prospect.

[0032] As can be seen, compared with the prior art, the present application has outstanding substantial characteristics and significant progress, and the beneficial effects of its implementation are also obvious. BRIEF DESCRIPTION OF DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0034] Figure 1 This is a flowchart of a high-voltage DC power supply method based on LSTM prediction provided by the present invention; Figure 2 This is a schematic diagram of a high-voltage DC power supply system based on LSTM prediction provided by the present invention. Figure 3 This is a schematic diagram of a data center HVDC power supply architecture based on an LSTM prediction-based high-voltage DC power supply method provided by the present invention. Figure 4 This is a schematic diagram of a segmented high-voltage DC bus structure for a data center based on an LSTM prediction-based high-voltage DC power supply method provided by the present invention.

[0035] The components include: 1. Electricity price prediction module; 2. Cost calculation module; 3. Power supply decision module; 4. Topology control module; 5. Plan generation module; 6. Parameter update module; 31. DC bus; 311. Segmented bus; 312. Bus tie switch; 313. DC / DC isolation module; 314. Data center power distribution cabinet; 315. Data center server; 316. Battery bank; 317. Supercapacitor bank; 318. AC / DC rectifier module; and 319. Transformer. Detailed Implementation

[0036] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following implementation methods.

[0037] Example 1: like Figure 1 As shown, this embodiment provides a high-voltage DC power supply method based on LSTM prediction, which includes the following steps: Step S1: Obtain historical electricity price data and current electricity price data, construct an LSTM model, use the LSTM model to predict the electricity price data for the target time period, and obtain the direct power supply cost for the next time slot; Step S2: Establish a joint objective function and calculate the power supply cost of each battery pack. The parameters of the joint objective function include at least one of the following: charging price, battery charging and discharging efficiency, bus power constraint, and energy storage life constraint. Step S3: Based on the direct power supply cost and the power supply cost of each battery pack, determine the power supply source and power allocation for the current time slot in the segmented high-voltage DC bus topology; Step S4: Control the AC rectification and DC isolation switching of the corresponding segments to establish or disconnect the energy channels between each battery pack and the bus, and generate a charging and discharging plan; Step S5: Update the forecast of direct power supply costs and the charge / discharge plan based on new observation data before reaching the next time slot.

[0038] By introducing the LSTM prediction model into the high-voltage direct current power supply system to dynamically predict the target period electricity price, the cost of different power supply paths is quantified and forward decision is realized, which can make the power supply control strategy have self-adaptability and time sequence foresight, and meet the needs of timely response of data center power supply cost, flexible energy storage scheduling and reliable prediction accuracy.

[0039] Specifically, by acquiring historical and current day electricity price data to construct an LSTM model, the electricity price prediction is no longer dependent on fixed assumptions and has long-term and short-term trend separation capability, providing continuity support in time dimension for power supply cost determination; by establishing a joint objective function, the charging electricity price, battery charging and discharging efficiency and energy storage life constraints are unified into the optimization target, so that the calculation of power supply cost not only reflects the economic factors, but also considers the safety of energy storage recycling and power boundary; by comparing the direct power supply cost and battery power supply cost in the segmented high-voltage direct current bus topology to determine the power supply source and power distribution, the system energy supply path can balance between economic optimization and capacity constraint; by controlling the on-off of AC rectification and DC isolation of each segment to establish or break the energy channel between the battery pack and the bus, real-time switching and coordinated operation between different energy units are realized; by rolling updating the prediction model and charging and discharging plan according to new observation data before the next time slot, the control instruction can continuously maintain dynamic consistency with market signals and operating state, so as to realize the unification of economic dispatching and high-reliability power supply, and meet the needs of low-cost, continuity and intelligent power supply of high-power density data centers.

[0040] Hereinafter, according to the embodiments of the present application, the above steps S1 to S5 are specifically described.

[0041] In step S1, historical electricity price data and current day electricity price data need to be acquired, and an LSTM model for electricity price prediction is constructed to realize dynamic prediction of electricity price in the target period and quantification of direct power supply cost in the next time slot. This step takes data centers as typical load scenarios, and aims to identify electricity price trend in advance to provide basis for subsequent energy management and power supply path decision.

[0042] Specifically, in the embodiments of the present application, the historical electricity price data can include market settlement price sequences in multiple time periods, and the current day electricity price data is real-time electricity price information of each time slot in the current day. At the same time, in order to ensure the continuity and consistency of time sequence, the above data can be uniformly sampled and time-aligned.

[0043] In some embodiments of the present application, during the model construction process, the LSTM model undertakes the core calculation function of electricity price prediction, the model input can include historical electricity price data and current day electricity price data, and can be extended to simultaneously input any one or more of distributed power output, real-time load of data center and state of charge of battery; the model structure contains input layer, several hidden layer units and output layer, the gating mechanism can be introduced in the hidden layer to realize the separated modeling of long-term trend and short-term fluctuation, the long-term gating is used to capture the trend change across days and weeks, and the short-term gating is used to describe the rapid fluctuation of peak electricity price segment, the next time slot prediction electricity price of the model output is used for subsequent cost calculation.

[0044] Among them, the distributed power output reflects the supply capacity of local power generation, the real-time load reflects the current energy consumption level of the data center, and the state of charge of the battery reflects the adjustable energy reserve. These features are aligned with the electricity price data sequence through a unified time index, can be input into the model in the form of a multi-dimensional tensor, and improve the joint identification ability of complex energy state and market signals; at the same time, in order to ensure the consistency of input dimensions, all feature data are normalized with the same scale.

[0045] In this step, by introducing multi-source operating parameters at the input end of the LSTM model, the prediction model can establish a dynamic mapping between electricity price and system operating state in a multi-dimensional data feature space, and the following technical effects can be achieved: First, the model receives historical electricity price data and current day electricity price data, and at the same time, the distributed power output, real-time load and state of charge of the battery are used as variable inputs, so that the prediction process takes into account the time series change and energy supply and demand characteristics, forming an overall response ability to the electricity price trend, thereby maintaining the prediction stability in the electricity price fluctuation scenario; Second, the gating mechanism separates the long-term trend and short-term fluctuation in the hidden layer output, so that the model maintains trend memory in long-term sequence, while having rapid adjustment ability under short-term disturbance, avoiding prediction deviation and cumulative error caused by single time scale modeling, and improving the adaptability of the prediction result to time series change; Third, the model realizes the time sequence dependence identification between the data center power supply related variables under the synergistic action of multi-input and gating mechanism, so that the prediction output can more accurately reflect the direction and fluctuation amplitude of the electricity price change in the future time slot, providing high-precision reference for subsequent power supply cost calculation and scheduling decision, and enhancing the response reliability and economy of the system in the dynamic market environment.

[0046] The predicted electricity price output by the model is mapped into the direct power supply cost of the next time slot through parameters such as rectifier efficiency and line loss, and the mapped cost is used for subsequent energy supply strategy selection and power distribution decision; at the same time, the prediction result can be compared and archived with the actual electricity price, so as to perform precision evaluation and model updating in the future.

[0047] For example, the input window length can be set to 96 time steps, corresponding to a one-day cycle at a 15-minute resolution; the number of hidden layer units can be 64 to balance computational overhead and feature representation capability; the electricity price input can be standardized in the [0,1] interval.

[0048] At this point, step S1 completes the collection, processing, and prediction of electricity price data, and obtains the direct power supply cost for the next time slot, providing basic data for subsequent energy allocation and optimization control.

[0049] In step S2, the core task is to establish a joint objective function to calculate the power supply cost of each battery pack. By combining parameters such as charging price, battery charging and discharging efficiency, bus power constraints, and energy storage life constraints, an optimization model that can quantify the power supply cost of batteries is constructed, providing a basis for subsequent energy allocation and scheduling decisions.

[0050] Specifically, in the embodiments of this application, the joint objective function aims to minimize the energy supply cost per unit time. It considers factors such as the discharge power of the battery, the electricity price signal, and the energy efficiency parameters. The charging electricity price cost of each battery pack consists of two parts: the direct cost corresponding to the electricity price factor and the implicit cost corresponding to energy loss and lifespan reduction. The output power of each battery pack is used as the decision variable. By finding the optimal solution of the joint objective function, a balance between cost and constraints is achieved.

[0051] In the cost modeling stage, the energy loss during the discharge process is determined by the battery's charging and discharging efficiency. Confirmed, when At the same time, the equivalent energy consumed during discharge increases, thus driving up the unit charging cost; simultaneously, if the battery pack is in a high-rate charge-discharge state for a long time, its lifespan will decrease with the number of cycles. Therefore, a lifespan decay factor is introduced into the joint objective function. The joint objective function is used to characterize the long-term economic losses caused by over-discharge. Thus, the joint objective function can be expressed as a comprehensive mapping relationship between cost and constraints. Its logic is to minimize the weighted power supply cost of battery operation while ensuring the balance between load power demand and bus power.

[0052] In embodiments of this application, the parameters of the joint objective function may further include at least one of battery charge / discharge efficiency, bus power constraint, and energy storage lifetime constraint. Each parameter is derived from real-time or historical monitoring data to ensure the model's responsiveness to changes in operating conditions. The charge / discharge efficiency is determined by the battery model or operating curve. The bus power constraint reflects the limitations of the data center's total load and upstream power supply capacity. The energy storage lifetime constraint is used to suppress high-rate discharge. For example, when the upper limit of the bus power is... At any given time slot, the sum of the battery output power and the direct power supply must not exceed [a certain value]. When the life constraint coefficient Taking the experience value 0.01-0.05, the life loss and energy cost can be balanced.

[0053] Further, to ensure the computability of the modeling process, the cost calculation of each battery pack is performed in time slices, the current electricity price, load demand and battery operating state are read in each prediction time slot, the discharge power of each battery pack is initially allocated, and then the target function is iteratively corrected, and the feasibility of the bus power constraint and the energy storage remaining capacity constraint is checked each time the iteration is performed. If any constraint is triggered, the power allocation is adjusted again until the condition is met. This process ensures that the solution of the joint objective function meets both energy balance and economy.

[0054] In some embodiments of the present application, the joint objective function can also be used to calculate the available energy of each battery pack, which is defined as the energy value that can be discharged at the current state of charge of the battery, affected by the lower limit of the state of charge, the discharge efficiency and the temperature correction coefficient. The model incorporates the available energy constraint into the optimization process when calculating, so that the sum of the available energy of each battery pack is not less than a preset critical value, thereby ensuring that the system still has sufficient energy reserve at peak load. This constraint can be realized by introducing a safety margin coefficient , ensuring that the total available energy , wherein is a preset critical threshold. Exemplarily, if the system includes three battery packs, two of which are used for main power support and one for peak shaving assistance, the model can ensure that the total energy reserve is not less than the preset critical threshold when optimizing by the available energy constraint.

[0055] In this step, by introducing the available energy constraint into the joint objective function, the energy state of the battery pack is combined with the power supply cost calculation process, so that the optimization decision has the capacity self-balancing ability and the energy storage redundancy management characteristics, and the following technical effects can be achieved: First, the joint objective function outputs the available energy of each battery pack while calculating the power supply cost, so that the cost optimization result is not only dependent on the electricity price prediction, but also bound with the actual capacity of the energy storage, thereby avoiding the problems of deep discharge and capacity imbalance caused by pure economic judgment, and improving the long-term operation stability of the energy storage unit; Second, by keeping the sum of the available energy of each battery pack not less than a preset critical value under the available energy constraint, the system always has a dispatchable energy storage redundancy during operation, which can maintain normal power supply path when the main power source fluctuates or the electricity price abnormally rises, realizing the parallel of power supply continuity and energy security; Third, the constraint condition is embedded as a hard constraint in the joint objective function solution, which makes the optimization process achieve a dynamic balance between global cost minimization and safety boundary of energy storage, avoids the life attenuation caused by frequent charging and discharging of a single battery group, and improves the stability of power scheduling and the balanced utilization rate of the energy storage group, so that the system maintains long-term consistency in power supply economy and reliability.

[0056] Further, to reduce local optimization problems, the optimization solution of the joint objective function can use gradient descent, quasi-Newton method or reinforcement learning-based policy search algorithm. The optimizer fine-tunes the discharge power of each battery group at each time slot, so that the gradient of the objective function gradually converges. When it is detected that the bus power is close to the constraint boundary, the constraint weight can be dynamically adjusted to prioritize continuous power supply.

[0057] In some other embodiments of the present application, the joint objective function can also take into account the performance differences of different types of battery groups. For example, lithium-ion batteries have high efficiency and long cycle life, but are more expensive; lead-carbon batteries are less expensive and have fast response speed, but have limited energy density; by introducing a weighting coefficient to differentiate different battery groups, the collaborative optimization of multiple types of energy storage units can be achieved.

[0058] So far, step S2 realizes the quantification and constraint coordination from the electricity price signal to the energy storage operation cost, so that the system maximizes the economy while ensuring power stability. The joint objective function constructed has scalability and differentiability, providing a continuous optimization basis for subsequent energy allocation and scheduling.

[0059] In step S3, based on the direct power supply cost obtained in step S1 and the power supply cost of each battery group obtained in step S2, the power supply source and power distribution scheme of the current time slot are determined. The core is to consider the cost, energy constraint and power continuity based on the actual topology of the segmented high-voltage DC bus, to realize the coordinated control of different power supply paths, so that the system completes the load power supply with the lowest comprehensive cost in the current time slot.

[0060] As shown in Figure 3 and Figure 4 In the embodiments of the present application, the high-voltage DC power supply system of the data center mainly includes a transformer 319, a plurality of AC / DC rectifier modules 318, a high-voltage DC bus 31, a plurality of segmented buses 311, a bus tie switch 312, a battery group 316, a super capacitor group 317, a DC / DC isolation module 313, a data center power distribution cabinet 314 and a data center server 315.

[0061] The high-voltage alternating current (such as 10 kV) output by the substation is stepped down to an alternating current voltage (such as 1 kV) suitable for the data center side by the transformer 319, and the stepped-down alternating current is rectified by the AC / DC rectifier module 318 to form a high-voltage direct current (such as 800 V) and is connected to the high-voltage direct current bus 31. The high-voltage direct current bus 31 is the main direct current bus node of the system, and is divided into a plurality of segmented buses 311 in the length direction. Each segmented bus 311 can be connected to or isolated from the direct current bus 31 by the bus coupler 312, so as to connect or separate the corresponding segmented bus 311 from the direct current bus 31 when power dispatching, expansion, maintenance or fault removal is required. At the same time, the adjacent segmented buses 311 can be electrically isolated by the bus coupler 312 to realize regional power supply autonomy and fault limitation in abnormal period.

[0062] The battery pack 316 and the super capacitor pack 317 matched with the capacity ratio (such as 10:1) of the battery pack 316 can be arranged on the direct current bus 31 or the segmented bus 311. The battery pack 316 is used to provide continuous power supply capability and can directly support the energy of the segmented bus 311 within the voltage allowable range to ensure the continuity of power supply and the stability of bus voltage. The super capacitor pack 317 is used to support short-time load and provide transient power support when the load suddenly increases or the rectifier side fluctuates, so as to balance the bus voltage. At the same time, an energy exchange device can be arranged between the battery pack 316 and the super capacitor pack 317 to float charge the battery pack 316 when the super capacitor pack 317 has excess energy, so as to improve the cycle efficiency of the whole energy storage.

[0063] Further, each segmented bus 311 is also connected to at least two groups of DC / DC isolation modules 313. The DC / DC isolation module 313 is used to realize the isolation and voltage matching between direct currents of different voltage levels, so that the server side can obtain a suitable operating voltage while the high-voltage direct current bus 31 remains stable. At the same time, the output end of the DC / DC isolation module 313 is connected to the data center power distribution cabinet 314, and the power is distributed to a plurality of server racks and load nodes of the data center server 315 through the data center power distribution cabinet 314, so as to realize the distribution of electric energy in the system.

[0064] In some embodiments of the present application, the high-voltage DC bus 31 and its segmented bus 311 adopt a modular structure design. When the power consumption scale of the data center increases, the capacity can be expanded by adding AC / DC rectifier modules 318, battery packs 316, super capacitor packs 317 or DC / DC isolation modules 313 in parallel to the segmented bus 311, without the need to modify the original system topology. For example, when the server deployment scale is expanded to 1.1 times or 1.2 times of the original capacity, the capacity expansion can be completed by simply connecting additional DC / DC isolation modules 313 and battery packs 316 to the corresponding segmented bus 311, thereby improving the scalability and evolution capability of the system.

[0065] At the beginning of the current time slot, the direct power supply cost and the battery pack power supply cost obtained in the previous step can be read first. The numerical relationship between the two is compared, and the power supply source is selected according to the cost minimization principle. If the direct power supply cost is lower than all battery pack power supply costs, it is determined that the in-network rectification link power supply within the time slot is the main power supply; if at least one battery pack power supply cost is lower than the direct power supply cost, the corresponding battery pack is selected as the main power supply source, and the rectification link is reserved as an auxiliary power supply path for rapid compensation of power fluctuations or bus voltage deviation.

[0066] On this basis, to avoid bus fluctuations caused by frequent switching, a small cost difference threshold can be set. When the cost difference between the two power supply paths is less than the threshold, the power supply mode of the previous time slot is maintained to improve system stability. For example, if the difference between the direct power supply cost and the lowest battery power supply cost is less than 0.02 yuan / kWh, the two costs can be considered equivalent, and the original power supply source is maintained.

[0067] In this step, the cost minimization principle is introduced into the power supply decision-making, and the comparison results of the direct power supply cost and the battery pack power supply cost are used as the basis to establish a dynamic decision-making mechanism, so that the power supply path selection has self-optimization characteristics and can achieve the following technical effects: First, the selection of the power supply source is based on the cost parameters of the current time slot, which quantitatively compares the economic efficiency of transformer rectification power supply and multiple battery power supply, so that the power supply path can be matched with the lowest cost scheme in real time, reducing unnecessary energy conversion loss and improving overall power supply efficiency; Second, under the action of the cost minimization principle, the charging and discharging behavior of different battery packs and the rectification power supply strategy form a complementary relationship, so that the system can automatically switch the power supply mode under the scenes of electricity price fluctuations and load changes, avoiding the redundant operation and power waste in traditional static scheduling, and improving the flexibility of system energy distribution; Third, the dynamic decision of power supply source enables the high-voltage DC bus in the segmented topology to realize rapid adjustment according to the equivalent cost of each path, takes into account power balance and economic optimization, maintains the optimal operating state when energy flows between multiple paths, and thus realizes real-time economic scheduling and high-reliability operation of the power supply system, and improves the overall energy utilization rate and the intelligent level of scheduling.

[0068] In the embodiments of the present application, after determining the power supply source, the power distribution phase is entered, the distribution weight is established based on the cost difference, so that the source with low cost bears a higher output share, and the source with high cost reduces the power supply proportion, so as to realize the minimization of the overall operating cost. The power distribution is simultaneously constrained by the bus topology, the available power of the battery, and the load distribution of each segment.

[0069] For example, the total load power of the current time slot is denoted as , the unit cost of direct power supply is , the unit power supply cost of the i-th battery pack is , and the cost difference of each power supply path is defined as:

[0070] When , it indicates that the i-th battery pack supplies power more economically, and its output should be increased; when , it indicates that direct power supply is more economical, and the battery should be kept in low power or standby state. Accordingly, the distribution ratio can be determined according to the cost difference:

[0071]

[0072] wherein is the load power of the i-th battery pack, is the cumulative value of the cost difference of all sources with better economy, and if only one battery pack meets the economic condition, the pack bears the main power supply, and the remaining sources are supplemented according to the remaining power capacity. Further, the topology characteristics of the segmented high-voltage DC bus also need to be considered in the power distribution process. The DC bus is usually divided into several segments, each segment is independently connected to the corresponding server array and energy storage branch, and the segments are maintained at the same potential through DC isolation or voltage equalization link. Therefore, power distribution not only determines the total output, but also determines the distribution ratio of each segment under the topology constraint.

[0073] In the bus power distribution, the load power of each segment can be expressed as

[0074] , and the total demand of the segment satisfies: ​​​

[0075] On this basis, the power distribution target of each segment is calculated according to the real-time load proportion of each segment, and then the target power distribution is mapped to the corresponding power supply source according to the connection relationship and the allowed output of each battery pack. For example, when a certain battery pack is connected only to the first and second segments, its distributed power is limited to the distribution within these two segments; if a segment is maintained or electrically isolated at this time, its power supply path is shielded and does not participate in the distribution.

[0076] In this step, a weight construction method based on cost difference is introduced in the power distribution link, and the power output is distributed in combination with the topological relationship of the segmented high-voltage DC bus, so that the power distribution process is changed from static control to dynamic economic dispatching, which can achieve the following technical effects: First, the cost difference is used to establish the distribution weight, so that the power distribution ratio is directly related to the economy of each power supply path, thereby realizing the optimal distribution of energy flow while maintaining the total power constant under different power supply modes, reducing the energy proportion of high-cost paths, and improving the overall energy efficiency of the system; Second, by combining the topological relationship of the segmented high-voltage DC bus to distribute power output, the segmented bus can maintain load balance under different operating states, avoiding voltage fluctuations and energy loss caused by centralized power supply or single bus overload, and realizing collaborative power supply between segments; Third, the dynamic update of the power distribution weight enables the system to adjust the energy output structure according to real-time cost changes, making the bus power flow have time sequence adjustability and spatial coordination, thereby ensuring balanced discharge of energy storage units and efficient use of rectifier paths, and maintaining stable, efficient and economic operation of the power supply network in a complex electricity price environment.

[0077] In some embodiments of the present application, on the basis of topological distribution, the feasibility of bus constraints and energy storage constraints can be further verified, the output of each power supply path should be less than its maximum allowed output; after distribution, the available energy of each battery pack still needs to meet the minimum energy storage requirement of the next time slot to prevent energy overflow; when the bus power approaches the upper limit of the capacity or the available energy of the battery is insufficient, the output of this source is automatically reduced and the proportion of other sources is adjusted to maintain power balance and constraint safety.

[0078] For example, if the current total load power is 1200 kW, the direct power supply cost is 0.85 yuan / kWh, and the power supply cost of the two battery packs is 0.8 yuan / kWh and 0.9 yuan / kWh respectively. ​​0.78 yuan / kWh and 0.82 yuan / kWh, respectively, two battery banks are selected as the main power supply source, and the power ratio is calculated to be about 7:3 according to the cost difference; if the first segmented load accounts for 40% of the total load and the second segmented load accounts for 60%, the power is further distributed to the corresponding bus according to the topology mapping; the super capacitor is used to provide short-time support during segmented voltage fluctuation, starting impact or power mutation.

[0079] For example, if the direct supply cost is lower than any battery bank, the rectified direct supply is maintained, and the charging sequence and duration of the battery are arranged during the low-price period according to the predicted curve, realizing "low-price charging and high-price charging"; if the direct supply cost is higher than all battery banks, the corresponding segment is closed and isolated, and the battery is connected to the grid for discharging, the load is preferentially borne by the lowest-cost group, and the insufficient part is supplemented by the second-lowest-cost group; if only an individual battery bank has a cost lower than the direct supply, only this group is connected to the grid for discharging, and the remaining groups are on standby and are recharged when the rectified capacity allows; in any case, the available power of each battery bank is not less than a set threshold value, to ensure that power supply can still be maintained for a period of time when the upper power supply fails.

[0080] At this point, through step S3, dynamic switching of power supply based on cost signals and segmented power distribution are realized in each time slot, so that the power supply mode of the high-voltage DC bus can be adaptively adjusted according to market electricity prices and energy storage states. The output result of this step is the power supply source set of the current time slot and the power distribution instructions of each source in each segment, which provides accurate input for the scheduling execution of the next stage.

[0081] In step S4, based on the running topology of the segmented high-voltage DC bus and the power distribution result of step S3, the control of AC rectification on-off and DC isolation on-off is executed, the energy channel between each battery bank and the bus is established or disconnected, and the corresponding charging and discharging plan is generated within the current and subsequent time slots.

[0082] Specifically, at the entrance of the on-off control, the running state, power reference and safety boundary of each segment need to be checked for consistency, which can include whether the segmented target power is less than the upper limit allowed by the segment, whether the source output capacity covers the reference value, whether the pre-charging condition is met, whether the bus voltage deviation is within the allowed bandwidth, etc. After checking, the on-off decision is entered; the on-off of the rectifier side is used to establish or release the AC rectification channel, and the isolation on-off of the DC side is used to establish or disconnect the energy channel between the segment and each battery bank. In order to reduce the impact, the on-off action can be executed in the order of "soft first and hard last", that is, pre-charging first, closing the isolation second, and releasing the current limit last; in reverse, when exiting, first reduce the power to the zero power window, then disconnect the isolation, and finally exit the rectification.

[0083] In some embodiments of the present application, the determination of the switching time is driven by both energy and price signals, and a delay confirmation strategy is adopted to avoid high-frequency jitter. Each battery pack uses the energy remaining rate as the first trigger, denoted as ; the electricity price signal uses the predicted change rate of the electricity price in the adjacent time slot as the second trigger, denoted as ; in any segment, only when the following linkage conditions are met, the switching from standby to grid connection or from grid connection to standby is triggered: first, the corresponding source of the segment obtains a positive power reference in step S3 and the interface capacity is sufficient; second, the trigger combination of energy and price reaches the set threshold; third, the delay confirmation timing reaches. The core of the delay confirmation strategy is to maintain the shortest confirmation time window for the trigger condition, and within the time window, if the predicted difference of the electricity price in the adjacent time slot is lower than the set threshold, the original on-off state is maintained without switching. In this way, the electrical topology can be maintained stable in the stage of small price signal disturbance, reducing unnecessary thermal stress and contactor wear.

[0084] In this step, the energy remaining rate and the electricity price change rate are combined as the on-off criterion, and the delay confirmation strategy is set to suppress short-term fluctuation triggering, so that the on-off control has threshold memory and switching lag characteristics, and the following technical effects can be achieved: First, the on-off switching is only triggered when the energy state and the price signal jointly reach the criterion, avoiding frequent transitions caused by single factor driving, maintaining the stable opening and closing relationship of the power channel, and reducing the conversion loss and thermal stress caused by the high-frequency action of the rectifier and isolation device; Second, the delay confirmation maintains the original on-off state in the scenario where the predicted difference in the adjacent time slot is small, so that the small prediction noise no longer causes topology shock, reduces the repeated climbing and falling of the bus voltage and current, and maintains the continuity of the segment load distribution and the coherence of the scheduling plan; Third, the participation of the energy remaining rate in the criterion makes different segments prefer to maintain the channel stable near the capacity boundary, reduces the shallow cycle true consumption of the energy storage unit caused by short-period switching, and makes the on-off rhythm consistent with the rolling plan, improving the predictability and economy of the whole power supply control in multi-time slot operation.

[0085] In the energy trigger logic, the priority condition for discharging grid connection is that the energy remaining rate of the group is higher than the limit threshold and the temperature rise and health status are allowed; the priority condition for charging grid connection is that the energy remaining rate of the group is lower than the energy supplement threshold and the price signal or the segment power reference allows charging supplement. The price trigger logic is used to accelerate or delay the switching: when the price rises significantly and the group bears the discharging power in step S3, the delay timing can be shortened; when the price falls significantly and the group is used for charging supplement, the delay timing can be shortened; in other cases, the nominal delay is maintained. Exemplarily, the discharging threshold and the energy supplement threshold of the energy remaining rate can take the target interval boundary given by the operation strategy.

[0086] Meanwhile, the delayed confirmation adopts a three-stage of "initiation-confirmation-execution": the initiation point is the time when the trigger combination is first met; the confirmation point is the first expiration checkpoint after the initiation; the execution point is the on-off action window entered after the confirmation. If the adjacent time slot price prediction difference is detected to be lower than the threshold at the confirmation point, the timer is cleared and the original state is maintained.

[0087] For the rectification channel and DC isolation, it follows the principle of segmented priority and cross-segment restriction at the execution level. The rectification channel of each segment is only set for the target power of the segment, and when the segment target power is zero and there is no cross-segment support task, the segment rectification remains off; the isolation channel of each battery pack is only closed in the segment it is connected to, and the power reference of the segment is the upper limit; when the same battery pack has connection conditions across two or more segments, it first meets the power reference of the segment, and then participates in adjacent segment support according to the remaining capacity; and all closing actions are performed after pre-charging is completed, and pre-charging is allowed to close after the pre-charging criterion is met, and the exit action is executed in the order of "power reduction-disconnection-precharge release-exit rectification", and before exiting, check whether the bus voltage falls into the dead zone.

[0088] In the embodiments of the present application, after the on-off control is completed, charge and discharge plans need to be generated for the current and several subsequent time slots in the rolling window. The plan generation takes the segmented power reference, the available power of each battery pack, the health state, and the environmental thermal boundary as input, and outputs the power reference trajectory at the time slot level and the grid / tie flag, and is accompanied by a rotation priority sequence, which is used to allocate the cycle consumption and thermal load in the case of multiple groups with equivalent economic efficiency, so as to delay the life degradation of individual battery packs.

[0089] Further, the calculation of the priority sequence combines the cycle number, the current temperature rise, and the state of charge in three dimensions to form a time slot application-oriented sorting result. The higher the cycle number, the later the sequence position; the higher the temperature rise, the later the sequence position; when the state of charge is in the high-risk zone (too high or too low), the later the sequence position; when the state of charge is in the target middle zone, the sequence position is in the front. Based on the sorting, in the equivalent cost and equivalent topology accessibility scenario, the battery pack with a higher sequence position is preferentially called to undertake the discharge or charge task of the current time slot, and the remaining battery packs participate according to the remaining power quota.

[0090] In this step, the rotation priority sequence is constructed based on the cycle number, temperature rise, and state of charge, and the power-on and standby order of each battery pack is arranged accordingly, so that the scheduling is changed from static fixation to adaptive rotation facing state quantities, which can achieve the following technical effects: First, the rotation priority sequence allows high-frequency units to give way to low-frequency units in time, the depth of discharge and cumulative cycles are distributed among units, excessive concentrated use is inhibited, capacity degradation is more balanced, standby capacity remains in a stable range, and energy redundancy can be called upon to maintain continuity of power supply and predictability of scheduling. Second, the temperature rise is used as a key input to make thermal constraints explicit, thermal loads are actively dispersed in rotation, the thermal gradient and hot spot formation of each section of the bus are suppressed, the thermal stress accumulation rate of insulation and conductor is reduced, the internal resistance changes more smoothly during charging and discharging, the power command response tends to be stable, the voltage drop and rise amplitude is reduced, and the frequency of thermal-related protection action triggering is reduced. Third, the state of charge is combined to evaluate the priority, so that different units alternate to undertake power output and charging tasks around the appropriate SOC interval, the deep and shallow cycle ratio is more reasonable, the equivalent loss during energy path switching is reduced, the power distribution and capacity boundary are kept in coordination, and the plan rolling adjustment under price and load disturbance is more smooth, and the comprehensive operating cost and available capacity utilization rate show stable income.

[0091] In some embodiments of the present application, to ensure the executability of the plan, the plan generation can include two layers of constraint checking. The first layer checks time consistency, that is, in any time slot within the window, the grid-connected state and energy channel on-off of the plan are not contradictory, and the planned power does not exceed the interface capacity and section upper limit. The second layer checks energy continuity, that is, the rolling energy balance does not cause the energy surplus rate to fall below the limit threshold or exceed the upper limit, and necessary control margin is reserved for the next time slot at the end of any time slot. If either check fails, adjust according to the "smooth first, then reduce" strategy: first smooth the power trajectory on the time axis, still not satisfied then evenly reduce, and in unavoidable scenarios, adjust the priority sequence to delay the task allocation of high-cycle and high-temperature battery packs.

[0092] At the same time, in order to maintain interface uniformity with the power distribution of step S3, the power reference of each section and each source in the plan can refer to the result of step S3 as an upper limit, and the command value of any source is not expanded in this step. When the delay confirmation trigger maintains the original state, the plan remains consistent with the previous time slot; when the confirmation trigger switches, the plan sets a "zero power transition section" in the switching window to complete the pre-charging and isolation closing; for cases that require cross-section support, the plan sets a transition section in both sections at the same time, and ensures that the voltage deviation monitored on both sides is lower than the allowed bandwidth before releasing the current limit. Exemplarily, a single switching window can include consecutive segments of pre-charging, closing and current limit release, and the total time is less than the minimum control period allowed by the section.

[0093] Further, in terms of safety and protection, the following bottom lines must be met during switching and plan execution: AC side fault or over-temperature warning takes priority, entering the forced exit branch and emptying the source's power plan in this time slot; DC side overcurrent, bus overvoltage or undervoltage triggers immediately derating and entering protection timing, and if it still does not recover after the timing expires, it goes into the exit process; before any grid closing action, the pre-charge current and bus voltage difference need to fall within the specified range, and if not, the closing is delayed until the window ends; after all protection actions are completed, the mismatched power and gap source of the current time slot are recorded by the monitoring side, providing a basis for subsequent statistics and model correction.

[0094] For the implementation of the delayed confirmation strategy, a threshold for the difference between the electricity prices of adjacent time slots can be set to determine whether to maintain the switching. The threshold can be configured as a fixed constant or updated adaptively according to the distribution of recent electricity price prediction errors; when the difference between adjacent time slots is less than the threshold and the current topology is stable, the existing switching state is maintained; when the difference exceeds the threshold and the energy and temperature rise conditions allow, the switching timing is entered; at the same time, the delay time window can be implemented in milliseconds to seconds at the device side and in time slot granularity at the scheduling side, and the two are superimposed to form a soft and hard double-layer debouncing. Exemplarily, the electricity price threshold for delayed confirmation can be of the order of magnitude of the root mean square error of recent price prediction.

[0095] Further, the generation of the rotation priority sequence can be completed inside the planner without exposing the weight details to the outside, and only the sorting and effective period are output. In time slots with equivalent cost and resource abundance, the battery groups at the back of the sequence automatically enter the "maintenance time slot" to charge at a small power to maintain the state of charge within the allowed interval and reduce the temperature rise; in time slots with resource shortage, the battery groups at the back of the sequence only participate under necessary conditions, and their participation power upper limit is lower than that of the groups at the front of the sequence to control the cumulative stress. Exemplarily, a number of maintenance time slots can be configured within a rolling window, and the battery groups with high cycle times and temperature rise indicators are preferentially arranged to enter the time period to gradually reduce their stress exposure. Exemplarily, the rotation priority sequence is updated at least once a day, and in periods of intense price fluctuations, it is rolled over and updated at a shorter period.

[0096] So far, step S4 converts the power allocation of step S3 into physically executable switching actions and time slot level charging and discharging trajectories, which not only restricts the switching frequency but also provides a rotation mechanism for the life and thermal management of multiple battery groups, thereby reducing the long-term cost and risk brought by invalid switching and unbalanced cycles while maintaining power supply continuity.

[0097] In step S5, the core task is to complete two types of updates before entering the next time slot: refresh the prediction results of the direct power supply cost based on new observation data; and perform rolling correction on the charging and discharging plan generated in the last time slot to make the power reference to be issued soon consistent with the price signal, load change and energy storage state.

[0098] Specifically, the data entry covers the latest daily power price sequence, distributed power output, real-time load, battery state of charge and temperature rise, rectifier and DC / DC efficiency estimation, online limit of segmented bus, and other observation quantities. All data are aligned under a unified timestamp, missing data is completed and abnormal data is weighted, peak spikes are retained but the training weight is reduced to ensure that short-term drastic changes can be identified by the model while avoiding excessive disturbance to the parameters. Then, the latest window data is used to drive the price predictor to obtain the price prediction of the next time slot and map it to the new direct power supply cost. If rectifier efficiency or line loss estimation changes during operation, the equivalent parameters are replaced in the cost mapping to maintain the consistency of the prediction and cost scopes.

[0099] In some embodiments of the present application, the model side can use a combination mechanism of incremental learning and sliding window. The window slides forward by one time slot, introduces the latest sample and discards the oldest sample, and performs small-step updates on the LSTM model weight. Abnormal sections participate in the update in a way of reducing the weight to avoid one-time extreme price pulling away from the long-term trend. In addition, when the price change rate exceeds the preset threshold, the prediction period can be adaptively shortened and the output frequency can be improved to reflect the market turning point faster. When the fluctuation falls, the prediction period returns to the regular granularity. The last version of the weight and the index are retained during the incremental learning process. If the verification error deteriorates beyond the limit, the stable version is immediately rolled back and the sample section of the failed update is recorded to prevent the spread of the misupdate. For example, when the price enters the rapid upward interval, the prediction period is shortened from the regular granularity to half the granularity.

[0100] In this step, by performing incremental learning of model parameters within the prediction period and combining the sliding window update mechanism, the LSTM model can continuously maintain the sensitivity and generalization ability to the price time series characteristics, adaptively adjust the prediction period in the volatile scenario, and achieve the following technical effects: First, the incremental learning mechanism updates only the local weight when receiving new observation data, retains the original structure stability, and thus maintains the convergence characteristics and gradually absorbs the latest price change information in the long-term operation, so as to continuously optimize the prediction accuracy over time and avoid performance degradation of the model caused by static parameters. Second, the sliding window mechanism ensures that the training samples are continuously rolling in time sequence, and the weights of new and old data are dynamically balanced, which prevents overfitting of recent abnormalities and captures the trend migration process, so that the prediction output forms a stable response between short-term fluctuations and long-term trends, and maintains the smoothness and reliability of the prediction curve. Third, the prediction period is adaptively shortened when the electricity price fluctuates sharply, so that the model update frequency matches the market change rhythm, reduces the decision deviation caused by prediction lag, and makes the power supply scheduling plan still timely and economical in the high-speed fluctuation interval, so as to realize the dynamic adaptability and continuous high-efficiency operation of the prediction model in the complex market environment.

[0101] The rolling correction on the planning side is to re-calculate the economic order of the next time slot by using the newly predicted direct power supply cost and the power supply cost of each battery pack, based on the source cost relationship given in step S3 and the established on-off state in step S4, but not directly triggering the on-off switching; only when the delay confirmation of step S4 has been met, and the energy remaining rate, temperature rise and interface capacity are all allowed, the change of grid-connected / off-grid flag is reflected in the planning layer; then, under the premise of not breaking the segmented limit and the upper limit of the interface, the power reference of each group is slightly redistributed, and the newly added economic advantage is preferentially allocated to the battery pack with lower cost and lower temperature rise; if the available power of any group is close to the safe lower limit, the discharge reference is reduced and the charging time slot is arranged in the plan to ensure that there is still sufficient reserve in the next time slot.

[0102] Further, the rotation priority sequence can also be evaluated synchronously at each rolling. The battery pack with higher cycle number, higher current temperature rise or state of charge in the edge interval is automatically moved backward to reduce the exposure to high stress; the battery pack in the target interval is moved forward to undertake the task of the current time slot. The new sequence formed only takes effect within the range of equivalent cost and equivalent topology accessibility, avoiding the reversal of the determined economic dominant relationship, and if the difference between the new prediction and the last version of the prediction is lower than the delay confirmation threshold, the plan remains unchanged, so that the power trajectory is continuous and the number of switching is controlled.

[0103] In some other embodiments of the present application, to ensure the consistency of execution, a time stamp and a signature identifier can also be implemented on the plan version, and any change is accompanied by the source, window range and rollback point. Two quick checks are carried out before the plan is implemented: energy continuity check to confirm that there will be no state of charge out-of-bound in the rolling window; and electrical safety check to confirm that the action sequence of pre-charging, current limiting, isolation and rectification can still be completed within the switching window. If the check fails, the plan is processed in the order of "priority smoothing, secondary derating, and necessary reservation of the original plan", and the mismatch information is written into the operation log for subsequent parameter setting.

[0104] For example, after the direct power supply cost is increased, the plan increases the discharge proportion of the battery pack with better economy without changing the confirmed on-off state, and arranges a maintenance time slot for the high-temperature-rise battery pack; if the subsequent two short-period predictions show that the price has fallen and the difference is again lower than the threshold, the regular period is restored, and the current plan is frozen to avoid repeated rewriting near the boundary.

[0105] Thus, through the processing in step S5, the direct power supply cost and charging / discharging plan are updated to be consistent before the start of the next time slot. This ensures both synchronization with new observation data and adherence to delayed confirmation and rotation mechanisms, guaranteeing a smooth plan, controllable switching, and sufficient reserves, providing a stable entry point for the next cycle.

[0106] In summary, this method, by establishing an electricity price prediction model, constructing a joint objective function, and executing power switching and power allocation under a segmented high-voltage DC bus topology, achieves coordinated operation of dynamic electricity price response, optimal energy storage cost allocation, and adaptive energy channel control. It can maintain system economy and power supply continuity under load fluctuations and electricity price changes, enabling DC power supply systems for data centers or high-power loads to have comprehensive optimization capabilities in energy management, including predictive driving, cost adaptation, and lifetime balancing.

[0107] It should be noted that, although the embodiments in this application are based on... Figure 1 Steps S1 to S5 are described sequentially, but this does not mean that steps S1 to S5 must be performed in a strict order. The reason this embodiment follows this order is... Figure 1 The order in which steps S1 to S5 are described is provided to facilitate understanding of the technical solutions of the embodiments of this application by those skilled in the art. In other words, in the embodiments of this application, the order of steps S1 to S5 can be appropriately adjusted according to actual needs.

[0108] Example 2: like Figure 2 As shown, this embodiment provides a high-voltage DC power supply system based on LSTM prediction, comprising: Electricity price prediction module 1 is used to build an LSTM model based on historical electricity price data and current electricity price data, and to predict the electricity price data for the target time period to obtain the direct power supply cost for the next time slot; Cost calculation module 2 is used to establish a joint objective function to calculate the power supply cost of each battery pack. The parameters of the joint objective function include at least one of the following: charging price, battery charging and discharging efficiency, bus power constraint, and energy storage lifetime constraint. Power supply decision module 3 is used to determine the power supply source and power allocation of the current time slot in the topology of the segmented high-voltage DC bus based on the direct power supply cost and the power supply cost of each battery pack. Topology control module 4 is used to control the AC rectification and DC isolation switching of the segmented high-voltage DC bus based on the output signal of power supply decision module 3, and to establish or disconnect the energy channel between each battery pack and the bus. The plan generation module 5 is used to generate and store the charging and discharging plan of the battery pack. The charging and discharging plan is updated synchronously with the power supply switching instructions. The parameter updating module 6 is configured to update the electricity price prediction module 1 and the charging and discharging plan based on new observation data before the next time slot arrives.

[0109] By means of the above technical scheme, the prediction driving and multi-source collaboration of power supply control are realized by arranging the electricity price prediction module 1, the cost calculation module 2, the power supply decision module 3, the topology control module 4, the plan generation module 5 and the parameter updating module 6 in the high-voltage direct-current power supply system, so that the optimal power supply path and efficient system operation can be maintained under the conditions of dynamic electricity price and complex load.

[0110] Specifically, the electricity price prediction module 1 uses historical and current electricity price data to construct an LSTM model, so that the system has the ability of time series learning and feedforward prediction of electricity price trend, and provides real-time basis for subsequent power supply cost determination; the cost calculation module 2 models the charging electricity price, battery efficiency, power constraint and life parameter by means of a joint objective function, so as to realize quantitative evaluation and economic judgment of power supply cost; the power supply decision module 3 selects the optimal power supply path and power distribution based on the cost comparison result, so that the energy scheduling process is converted from fixed logic to real-time optimization; the topology control module 4 executes on-off instructions on the AC rectifier and DC isolation channel according to the decision signal, so that the energy flow direction of each battery pack is automatically reconstructed according to the scheduling result, and the power transmission path and operation strategy are consistent; the plan generation module 5 records the charging and discharging sequence and keeps synchronization with the power supply switching instruction, so that the energy storage behavior is coordinated with the market price rhythm; and the parameter updating module 6 performs rolling correction on the model parameters and scheduling plan before each time slot, so that the system continuously maintains the prediction accuracy and scheduling consistency in long-term operation, thereby realizing the comprehensive optimization of cost, efficiency and stability of the high-voltage direct-current power supply of the data center.

[0111] In the embodiment, the power supply decision module 3 can include a power distribution unit configured to calculate a distribution weight according to the difference between the direct power supply cost and the power supply cost of each battery pack, and generate a power distribution instruction according to the distribution weight.

[0112] In the power supply decision module 3, the weight calculation mechanism based on the cost difference is introduced in the power distribution unit, so that the generation of the power instruction takes into account the economic efficiency and electrical constraint coordination, and the following technical effects can be achieved: First, the power distribution unit establishes a weight matrix for multiple power supply paths by taking the cost difference as the decision quantity, so that the output power of each path is directly linked to its economic contribution, thereby automatically reducing the high-cost channel and automatically increasing the low-cost channel, forming a power supply distribution process that dynamically self-balances with market signals, and reducing redundant energy flow and unnecessary power repeated switching. Second, the weight generation mechanism enables the power allocation result to be synchronized with the bus topology structure, and the power output proportion between different battery groups is adjusted in real time according to the respective capacity, efficiency and bus constraints, thereby avoiding uneven energy efficiency caused by overloading or idling of a single battery group, and improving the overall power utilization rate and bus operation stability of the system; Third, the power allocation instruction is uniformly issued by the power supply decision module 3 after real-time calculation, so that the system can quickly respond and reconstruct the power allocation pattern under the condition of sudden load change or sharp price fluctuation, maintain continuous output and voltage balance, and improve the adaptability and sustained economic operation ability of the high-voltage direct-current power supply system under complex operating conditions.

[0113] In summary, the system integrates the price prediction module 1, the cost calculation module 2, the power supply decision module 3, the topology control module 4, the plan generation module 5 and the parameter updating module 6 and other collaborative modules in the high-voltage direct-current power supply architecture, realizes the whole-link adaptive control from price prediction to power allocation, dynamically optimizes the power supply path in the multi-power parallel and price fluctuation environment, takes into account the power supply economy and operation stability, and ensures that the data center realizes efficient, low-cost and intelligent continuous power supply under the high power density scenario.

[0114] The above disclosure is only the preferred embodiment of the present application, but the present application is not limited thereto, any non-creative changes that can be thought of by those skilled in the art, and several improvements and refinements made without departing from the principles of the present application, should fall within the scope of protection of the present application.

Claims

1. A high voltage direct current power supply method based on LSTM prediction, characterized in that, The method comprises the following steps: obtaining historical electricity price data and current-day electricity price data, constructing an LSTM model, using the LSTM model to predict the electricity price data of a target period, and obtaining the direct power supply cost of the next time slot; establishing a joint objective function to calculate the power supply cost of each battery pack, the parameters of the joint objective function including at least one of the charging electricity price, the charging and discharging efficiency of the battery, the bus power constraint, and the energy storage life constraint; based on the direct power supply cost and the power supply cost of each battery pack, determining the power supply source and power distribution of the current time slot in the topology of the segmented high-voltage DC bus; controlling the on-off of the AC rectifier and the on-off of the DC isolation corresponding to the segment, establishing or breaking the energy channel between each battery pack and the bus, and generating a charging and discharging plan; updating the prediction of the direct power supply cost and the charging and discharging plan based on new observation data before reaching the next time slot.

2. The high voltage direct current power supply method based on LSTM prediction of claim 1, wherein, The input of the LSTM model includes at least one of historical electricity price data, current-day electricity price data, distributed power output, real-time load of the data center, and state of charge of the battery, and the LSTM model introduces a gating mechanism in the hidden layer output to model long-term trends and short-term fluctuations respectively. 3.The high voltage direct current power supply method based on LSTM prediction of claim 1, wherein, The joint objective function is also used to calculate the available power of each battery pack, and the sum of the available power of each battery pack is kept not lower than a preset threshold under the constraint of the available power.

4. The high voltage direct current power supply method based on LSTM prediction of claim 1, wherein, The determination of the power supply source is based on the comparison result of the direct power supply cost and the power supply cost of each battery pack, and the power supply source is selected according to the cost minimization principle in the current time slot.

5. The high voltage direct current power supply method based on LSTM prediction of claim 4, wherein, The determination of the power distribution establishes a distribution weight according to the difference between the direct power supply cost and the power supply cost of each battery pack, and distributes the power output according to the topology relationship of the segmented high-voltage DC bus.

6. The high voltage direct current power supply method based on LSTM prediction of claim 1, wherein, The control of the on-off of the AC rectifier and the on-off of the DC isolation corresponding to the segment includes: determining the switching time according to the energy remaining rate and the electricity price change rate of each segment of the battery pack, and the switching time is determined by using a delayed confirmation strategy, and the on-off state is kept unchanged when the difference between the adjacent time slot electricity price prediction is lower than a set threshold.

7. The high voltage direct current power supply method based on LSTM prediction of claim 1, wherein, The generation of the charging and discharging plan includes setting a rotation priority sequence for each battery pack, and the rotation priority sequence is determined according to the cycle number, current temperature rise, and state of charge of the battery pack. 8.The high voltage direct current power supply method based on LSTM prediction of claim 1, wherein, The update is performed by performing incremental learning on the LSTM model parameters in each prediction period, using a sliding window mechanism to continuously update the model weight, and adaptively shortening the prediction period when the electricity price fluctuates sharply.

9. A high voltage direct current power supply system based on LSTM prediction, characterized in that, It comprises: An electricity price prediction module is configured to construct an LSTM model based on historical electricity price data and current-day electricity price data, and to predict the electricity price data of a target period to obtain the direct power supply cost of the next time slot; A cost calculation module is configured to establish a joint objective function to calculate the power supply cost of each battery pack, and the parameters of the joint objective function include at least one of the charging electricity price, the charging and discharging efficiency of the battery, the bus power constraint, and the energy storage life constraint. a power supply decision module configured to determine a power supply source and power distribution of a current time slot in a topology of the segmented HVDC bus according to the direct power supply cost and the power supply cost of each battery pack; a topology control module configured to control AC rectification and DC isolation of the segmented HVDC bus according to an output signal of the power supply decision module, and to establish or disconnect an energy channel between each battery pack and the bus; a plan generation module configured to generate and store a charge-discharge plan of the battery pack, and to update the charge-discharge plan synchronously with a switching instruction of the power supply source; a parameter update module configured to update the electricity price prediction module and the charge-discharge plan based on new observation data before a next time slot arrives.

10. The high voltage direct current power supply system based on LSTM prediction of claim 9, wherein, The power supply decision module comprises a power distribution unit configured to calculate a distribution weight according to a difference between the direct power supply cost and the power supply cost of each battery pack, and to generate a power distribution instruction according to the distribution weight.

Images