Intelligent campus shared energy and computing power joint optimization scheduling method considering stepped carbon trading

Abstract

The application discloses a kind of wisdom campus shared energy storage-power calculation joint optimization scheduling method considering stepped carbon trading, obtains the relevant parameters of campus integrated energy system;State space and action space are constructed based on Q-learning algorithm, joint action is selected by ε-greedy strategy and multi-round training is carried out, and energy storage target power and power calculation target power are output;Based on YALMIP, a mixed integer linear programming model is constructed, the target power output by the upper layer strategy generation module is guided as a soft constraint, the minimum annual total cost of the system is taken as the target, and the optimal capacity configuration of shared energy storage and the operating power of each device in each period are solved;Energy storage configuration results, air conditioning operation and temperature change in each functional area, power calculation center scheduling curve and carbon trading allocation results are output and displayed.The present application aims to solve the technical problems of insufficient energy storage and power calculation cooperation, weak guiding of carbon trading mechanism and inaccurate multi-energy flow coupling modeling in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of smart grid and integrated energy system technology, and specifically relates to a smart campus shared energy storage-computing power joint optimization scheduling method that takes into account tiered carbon trading. Background Technology

[0002] With the deepening implementation of the "dual carbon" goals, the low-carbon and efficient operation of the integrated energy system of smart campuses, as complex energy-consuming units integrating teaching, research, and living, has become a research hotspot. Campus integrated energy systems typically include distributed photovoltaics, wind power, conventional electrical loads, air conditioning systems, and emerging data centers (computing centers), exhibiting complex characteristics of multi-energy coupling involving electricity, heat, and computing power. How to effectively integrate these resources to achieve economical, low-carbon, and efficient coordinated operation is a pressing technical problem that needs to be solved.

[0003] Existing technologies for the optimal scheduling of integrated energy systems have the following main shortcomings:

[0004] First, there is insufficient coordination between energy storage and computing load. Existing scheduling methods mostly manage shared energy storage power stations as independent flexible resources, and the coordination between their charging and discharging strategies and fluctuating renewable energy sources (photovoltaics, wind power) and various types of loads (teaching, dormitories, research) on campus needs improvement. Meanwhile, computing centers (data centers), as emerging "computing loads," possess flexible characteristics of time-shiftable and adjustable IT equipment power, but existing technologies typically treat them as rigid loads or only perform simple peak shaving and valley filling, failing to treat them as flexible and adjustable "virtual energy storage" resources for joint scheduling with physical energy storage. This limits the overall economic efficiency of the system and its ability to absorb renewable energy. Second, the carbon trading mechanism has weak guidance. Existing carbon trading models mostly use fixed carbon prices or simple linear ladders, failing to accurately depict the tiered growth characteristics of carbon emission rights trading prices, making it difficult to effectively incentivize the system to further explore emission reduction potential after exceeding free allowances. Meanwhile, existing models fail to quantify and incorporate the indirect emission reduction benefits of shifting computing load during the time period (such as transferring electricity consumption from high-carbon emission periods to low-carbon emission periods) into the optimization objectives, resulting in a lack of targeted decision-making in the system's pursuit of low-carbon operation. Thirdly, multi-energy flow coupling modeling is inaccurate. The thermal dynamic characteristics of buildings in different functional areas of the campus (such as teaching areas, dormitory areas, and research areas) vary and influence each other. Existing technologies often simplify the thermal dynamics of each building into independent models or ignore their coupling relationships, making it difficult to achieve precise control of air conditioning load. Furthermore, the efficient operation of computing centers relies on strict room temperature control, and their cooling load is closely coupled with the power of IT equipment. Existing technologies fail to uniformly model and optimize this deep coupling relationship of "electricity-heat-computing power," resulting in the inability to achieve optimal global energy allocation while ensuring user comfort and computing service quality.

[0005] In summary, existing technologies lack a unified optimization scheduling scheme that can comprehensively consider the coordinated scheduling of shared energy storage and computing power loads, the precise guidance of tiered carbon trading, and the multi-energy flow coupling modeling of "electricity-heat-computing power". This case arises from this. Summary of the Invention

[0006] The purpose of this invention is to provide a smart campus shared energy storage-computing power joint optimization scheduling method that takes into account tiered carbon trading, aiming to solve technical problems in the existing technology such as insufficient coordination between energy storage and computing power, weak guidance of carbon trading mechanism, and inaccurate multi-energy flow coupling modeling.

[0007] To achieve the above objectives, the solution of the present invention is:

[0008] A joint optimization scheduling method for shared energy storage and computing power in smart campuses, taking into account tiered carbon trading, includes:

[0009] Obtain operating parameters of the campus integrated energy system, building thermal dynamic parameters, computing center parameters, and tiered carbon trading parameters;

[0010] A two-layer optimization scheduling system is constructed, consisting of an upper-layer policy generation module and a lower-layer optimization solution module. In the upper-layer policy generation module, the state space and action space are constructed based on the Q-learning algorithm, the joint action is selected through the ε-greedy policy and multiple rounds of training are performed, and the target power of energy storage and the target power of computing power are output based on the acquired parameters.

[0011] In the lower-level optimization solution module, a mixed-integer linear programming model is constructed based on YALMIP. Guided by the target power output by the upper-level strategy generation module, and with the goal of minimizing the annualized total cost of the system, the optimal capacity configuration of shared energy storage and the operating power of each device in each time period are solved.

[0012] Output and display the energy storage configuration results, the air conditioning operation and temperature changes in each functional area, the computing center scheduling curve, and the carbon trading allocation results.

[0013] In the aforementioned upper-level strategy generation module, the state space includes four dimensions: energy storage charge status level, electricity price time period type, total output level of new energy, and benchmark load level of computing center; the action space is composed of energy storage actions and computing power actions. Energy storage actions include three modes: charging, discharging, and holding, while computing power actions include three modes: full load, low load, and delay.

[0014] In the aforementioned upper-level strategy generation module, an ε-greedy strategy is used for action selection. The exploration rate decreases as the number of training rounds increases, and an immediate reward is calculated after each action is executed in each time period. The immediate reward includes electricity purchase cost, carbon trading cost, computing power latency penalty, computing power time shift penalty, and SOC out-of-bounds penalty.

[0015] In the aforementioned lower-level optimization solution module, the annualized total system cost includes the investment and operation and maintenance costs of shared energy storage, the investment and operation and maintenance costs of computing power centers, the cost of time-of-use electricity purchases, the transaction costs of shared energy storage with various functional areas and computing power centers, the cost of computing power latency and time shift penalties, and the cost of tiered carbon trading.

[0016] The constraints in the aforementioned lower-level optimization solution module include shared energy storage charge state evolution and rate constraints, building thermal dynamic equation constraints for each functional area, indoor temperature comfort range constraints, air conditioning power ramping constraints, upper and lower limits and ramping constraints for IT equipment power in the computing center, daily computing power conservation constraints, minimum following ratio constraints during working hours, power balance constraints, and tiered carbon trading constraints.

[0017] The aforementioned building thermal dynamic equation constraints are constructed based on an equivalent thermal parameter model, including the thermal balance relationship between wall heat capacity, wall thermal resistance, solar radiation, indoor and outdoor temperature difference, and air conditioning heating / cooling power.

[0018] The aforementioned tiered carbon trading parameters include the carbon trading base price, price growth rate, segment interval length, grid purchase carbon emission quota factor, and actual carbon emission factor. The lower-level optimization solution module decomposes net carbon emissions into multiple tiers by introducing positive and negative carbon emission variables and segment variables, and calculates carbon trading costs using incremental carbon trading prices.

[0019] The aforementioned lower-level optimization solution module is also used to quantify the emission reduction benefits of computing power time shift. By calculating the difference in the grid carbon emission factor during peak and valley periods, the carbon emission reduction caused by the shift of computing power load from peak to valley periods is included in the carbon emission reduction.

[0020] The target power of energy storage and the target power of computing power output by the above-mentioned upper-level strategy generation module serve as soft constraints as guidance. By introducing slack variables, deviations are penalized to avoid infeasibility conflicts between hard constraints and physical constraints.

[0021] By adopting the above solution, the present invention has at least the following technical effects or advantages:

[0022] (1) This invention actively balances the "breadth of policy exploration" and "precision of physical constraints" in the algorithm dimension through multi-round training of upper-layer Q-learning, precise solution of lower-layer MILP, and progressive collaborative logic guided by soft constraint relaxation. This changes the static mode in traditional technology where reinforcement learning is prone to producing physically infeasible solutions or dimensional explosion of pure mathematical programming solutions. It can effectively avoid the model conflict problem caused by directly using AI output as hard constraints, and realize more efficient and robust generation and engineering implementation of scheduling strategies for complex multi-energy systems.

[0023] (2) This invention not only schedules energy storage and computing power as independent resources separately, but also automatically exploits the flexible characteristics of IT equipment tasks that can be shifted and power that can be adjusted, equating them to "virtual energy storage" and physical shared energy storage for joint optimization. Based on time-of-use electricity prices and carbon price signals, the system autonomously executes the coordinated action of "valley storage and peak release, valley loading and peak delay," and accurately quantifies the indirect emission reduction benefits brought about by load shifting under the premise of strictly meeting the daily total computing power conservation and working period following ratio. This integrated autonomous control capability of "physical flexible resources + virtual computing power resources" significantly broadens the adjustment boundary of the campus microgrid and improves the system's adaptive absorption level of fluctuating renewable energy.

[0024] (3) This invention enables the optimization system to accurately identify the thermal inertia differences and load characteristics of different functional areas (teaching, research, and dormitory) in a multi-dimensional topology by constructing building thermal dynamic constraints and zoning comfort zone control based on the equivalent thermal parameter (ETP) model. By deeply coupling wall thermal capacity / thermal resistance, solar radiation heat gain, air conditioning power ramping and computing center PUE cooling, the abstract thermodynamic equations are transformed into linear constraints that can be solved efficiently, which greatly reduces the computational complexity of multi-energy flow coupling modeling and provides refined decision support from "macroscopic power balance" to "microscopic indoor thermal comfort and computer room temperature control guarantee".

[0025] (4) This invention uses the nonlinear growth characteristics of tiered carbon trading prices as a driving switch to trigger deep emission reduction behavior in the system, realizing the underlying linkage between power economic operation and carbon emission management. By introducing positive and negative carbon emission decomposition variables and a segmented incremental carbon price calculation model, and combining the difference in grid carbon emission factors during peak and valley periods, the high carbon emissions avoided by the spatiotemporal transfer of computing power load are accurately included in the carbon emission reduction. This makes the dispatching entity face a step-by-step increase in marginal emission reduction costs after exceeding the free quota, thus actively transferring energy flow and computing power flow to low-carbon periods, ultimately breaking down the barriers between economics and environmental protection at the decision-making level, and achieving cross-objective synergy between "precise guidance of tiered carbon prices" and "computing power-energy storage joint carbon reduction". Attached Figure Description

[0026] Figure 1 This is a schematic diagram of a smart campus shared energy storage-computing power joint optimization scheduling method that takes into account tiered carbon trading, according to the present invention.

[0027] Figure 2 Here are four typical daily outdoor temperature curves;

[0028] Figure 3 The graphs show the wind load curves for four typical sunlit teaching areas.

[0029] Figure 4 The following are wind and solar load curves for four typical sunlit scientific research and experimental areas;

[0030] Figure 5 Here are the solar load curves for four typical residential areas under the sun;

[0031] Figure 6 A graph showing the operating power of the air conditioners in the teaching area and the changes in indoor temperature.

[0032] Figure 7 A graph showing the operating power of the air conditioner and the change in indoor temperature in the scientific research and experimental area;

[0033] Figure 8 A graph showing the operating power of air conditioners in the dormitory area and the changes in indoor temperature.

[0034] Figure 9 This is a typical daily scheduling curve for computing power centers in spring.

[0035] Figure 10 This is a typical daily scheduling curve for computing power centers during summer.

[0036] Figure 11 This is a typical daily scheduling curve for computing power centers in autumn.

[0037] Figure 12 This is a typical daily scheduling curve for computing power centers during winter.

[0038] Figure 13 This is a graph showing the time-shifting effect of computing load.

[0039] Figure 14 A graph showing the power consumption of electricity purchased on a typical day on campus versus the charging and discharging power of shared energy storage.

[0040] Figure 15 The convergence curve for Q-learning training. Detailed Implementation

[0041] This invention provides a smart campus shared energy storage-computing power joint optimization scheduling method considering tiered carbon trading, comprising:

[0042] Obtain operating parameters of the campus integrated energy system, building thermal dynamic parameters, computing center parameters, and tiered carbon trading parameters;

[0043] A two-layer optimization scheduling system is constructed, consisting of an upper-layer policy generation module and a lower-layer optimization solution module. In the upper-layer policy generation module, the state space and action space are constructed based on the Q-learning algorithm, the joint action is selected through the ε-greedy policy and multiple rounds of training are performed, and the target power of energy storage and the target power of computing power are output based on the acquired parameters.

[0044] In the lower-level optimization solution module, a mixed-integer linear programming model is constructed based on YALMIP. Guided by the target power output by the upper-level strategy generation module, and with the goal of minimizing the annualized total cost of the system, the optimal capacity configuration of shared energy storage and the operating power of each device in each time period are solved.

[0045] Output and display the energy storage configuration results, the air conditioning operation and temperature changes in each functional area, the computing center scheduling curve, and the carbon trading allocation results.

[0046] The following is a detailed explanation.

[0047] (I) System Overall Architecture

[0048] The two-layer optimization scheduling system constructed in this invention includes an upper-layer policy generation module and a lower-layer optimization solution module. The upper layer adopts the Q-learning algorithm to learn the optimal decision pattern through offline training and generate the joint scheduling target power of shared energy storage and computing center. The lower layer is based on mixed-integer linear programming, guided by the target power output by the upper layer as soft constraints. Under the premise of satisfying the physical constraints of the system, it solves the optimal capacity configuration of shared energy storage and the operating power of each device in each time period, thereby minimizing the annualized total cost of the system.

[0049] (ii) Upper-level Q-learning strategy generation model

[0050] The upper-level policy generation module models the system operation as a Markov decision process, learning the optimal scheduling policy through the interaction between the agent and the environment.

[0051] The state space is defined as a four-dimensional combinatorial state:

[0052]

[0053] in, Let t be the state vector for time period t; Discretization level for shared energy storage state of charge; This indicates the electricity price time period type, with values ​​of 1, 2, and 3 corresponding to off-peak, average, and peak periods, respectively. The total output level of new energy sources is categorized into three levels: 1, 2, and 3, which correspond to low, medium, and high levels, respectively. This represents the baseline load level of the computing center, with values ​​of 1, 2, and 3 corresponding to low, medium, and high levels, respectively.

[0054] The action space consists of a combination of energy storage actions and computing power actions:

[0055]

[0056] in, For the joint actions during time period t; For energy storage, values ​​1, 2, and 3 correspond to the three modes of charging, holding, and discharging, respectively. This represents the computing power action, with values ​​of 1, 2, and 3 corresponding to three modes: full load, low load, and latency, respectively.

[0057] Energy storage actions are mapped to charge / discharge power commands:

[0058]

[0059] in, t represents the target energy storage power during time period t, with positive values ​​indicating discharge and negative values ​​indicating charging; This is the energy storage action step size coefficient, with a value between 0.2 and 0.4. Rated power for shared energy storage.

[0060] Computing power actions are mapped to IT equipment power target values:

[0061]

[0062] in, The target power of IT equipment in the computing center during time period t; and These are the maximum and minimum power of the IT equipment, respectively. Let t be the computing power required during time period t.

[0063] The reward function is used to evaluate the immediate benefit after performing an action, and is defined as:

[0064]

[0065] in, The instant reward for time period t; For electricity purchase costs; For carbon trading costs; The cost of computing power latency penalty; The cost of time-shifting penalty for computing power; The cost of penalties for SOC overstepping boundaries; , , , , These are the normalized scaling factors for the corresponding costs, used to balance the magnitudes of each cost item.

[0066] Electricity purchase cost The calculation is as follows:

[0067]

[0068] in, The time-of-use electricity price for period t; This refers to the net load after deducting the output of new energy sources.

[0069] Carbon trading costs Adopting a tiered carbon trading model:

[0070]

[0071] in, The tiered carbon price for time period t; t represents the net carbon emissions during the period t.

[0072] Computing latency penalty cost and time-shift penalty costs They are respectively:

[0073]

[0074]

[0075] in, This is the latency penalty coefficient for computing power; This is the computing power time-shift penalty coefficient; The original computing power requirement for time period t; This represents the actual power of the IT equipment.

[0076] SOC boundary violation penalty cost for:

[0077]

[0078] in, This is the SOC (System Oscillator) out-of-bounds penalty coefficient; The energy storage state of charge during time period t; and These are the minimum and maximum values ​​of SOC, respectively.

[0079] The Q-learning algorithm updates the Q value using the Bellman equation:

[0080]

[0081] in, For state Take action below Action value function; The learning rate ranges from 0.01 to 0.1. This is a discount factor, with a value range of 0.9 to 0.99; The state for the next time period; This represents the maximum Q value in the next state.

[0082] Employing an ε-greedy strategy to balance exploration and exploitation:

[0083]

[0084] in, For state The maximum Q value is given; ε is the exploration rate, which decays exponentially with the number of training epochs.

[0085]

[0086] In the formula, This is the initial exploration rate; Minimum exploration rate; To explore the rate of decay factor; This represents the current training round number.

[0087] (III) Lower-level optimization solution model

[0088] The lower-level optimization solution module is built on mixed integer linear programming, with the objective function being the minimum annualized total cost of the system, and constraints such as shared energy storage, building thermal dynamics, computing centers, power balance, and tiered carbon trading.

[0089] The system's annualized total cost includes two parts: investment and maintenance costs, and operating costs.

[0090]

[0091] in, This represents the system's annualized total cost; This represents the annualized investment and maintenance costs. This represents the annual operating cost.

[0092] The investment and operation costs are the equivalent annual value of the shared energy storage and computing center.

[0093]

[0094] In the formula, The unit power investment cost for shared energy storage, expressed in yuan / kW; The unit capacity investment cost for shared energy storage, expressed in yuan / kWh; The unit operation and maintenance cost of shared energy storage is expressed in yuan / (year·kW); and These are the rated power and rated capacity of the energy storage, respectively. This refers to the number of days in the energy storage lifespan. , These represent the unit power investment cost and operation and maintenance cost of the computing center, respectively. Maximum power of IT equipment This refers to the lifespan of the computing center in days.

[0095] Operating costs include electricity purchase costs, energy storage trading costs, energy storage service fees, computing power penalty costs, and carbon trading costs.

[0096]

[0097] In the formula, The number of days for each typical day. This represents the number of typical days; The typical daily electricity purchase cost; The energy storage transaction cost for a typical day; The energy storage service fee is for a typical day. The computing power penalty cost for a typical day; This represents the typical carbon trading cost for a day.

[0098] Electricity purchase cost This includes the electricity costs for each functional area and the computing center:

[0099]

[0100] in, The number of time periods per day is set to 24. The number of functional zones on campus; The time-of-use electricity price for period t; The power purchased by functional area n during a typical day's time period t; This represents the electricity purchased by the computing center during a typical daily time period t.

[0101] Energy storage transaction costs The cost of trading electricity between shared energy storage and various functional areas and computing centers:

[0102]

[0103] In the formula, The electricity price for energy storage transactions during time period t. The trading power between functional zone n and energy storage, with positive values ​​indicating discharge and negative values ​​indicating charging; This refers to the trading power between computing centers and energy storage.

[0104] Energy storage service fee Service fees charged to each functional area for shared energy storage:

[0105]

[0106] In the formula, This is the unit price for the service fee. For energy storage discharge power, The absolute value of the energy storage charging power satisfies and .

[0107] Computing power penalty cost Includes delay penalties and time-shift penalties:

[0108]

[0109] In the formula, This is the latency penalty coefficient for computing power. This is the computing power time shift penalty coefficient. The computing power requirement for time period t. This represents the actual power of the IT equipment.

[0110] (1) Shared energy storage constraints

[0111] Equation of the state of charge:

[0112]

[0113] in, This represents the energy storage capacity during a typical daily time period (t). Self-discharge efficiency; and These are charging efficiency and discharging efficiency, respectively. and These are charging power and discharging power, respectively. The time interval is 1 hour.

[0114] Charge / discharge power and capacity rate constraints:

[0115]

[0116]

[0117] In the formula, Energy multiplier This is a binary variable representing the charging / discharging state (1 indicates discharging, 0 indicates charging).

[0118] SOC upper and lower bound constraints:

[0119]

[0120] in, and These are the minimum and maximum values ​​of SOC, respectively.

[0121] Initial and final SOC balance constraints:

[0122]

[0123] in, This represents the initial state of charge of the energy storage.

[0124] (2) Building thermal dynamic constraints

[0125] Based on the equivalent thermal parameter model, the wall temperature evolution equation is as follows:

[0126]

[0127] in, Let be the heat capacity of the i-th wall, in Wh / ℃; Let be the temperature of the i-th wall during a typical daytime period t. Let be the thermal resistance of the i-th wall, in °C / W; Indoor temperature; Outdoor temperature; Let represent the solar radiation heat gain of the i-th wall, expressed in W.

[0128] Indoor temperature evolution equation:

[0129]

[0130] in, The equivalent heat capacity of the room is expressed in Wh / ℃. The heat gain power of the window from solar radiation, in W; This refers to the air conditioner's thermal power, measured in W.

[0131] Air conditioning operation constraints:

[0132]

[0133]

[0134]

[0135] in, and These are heating power and cooling power, respectively; COP is the air conditioner's energy efficiency ratio. This is the upper limit of the air conditioner's electrical power. This is a binary variable representing the heating / cooling mode; a value of 1 indicates heating, and a value of 0 indicates cooling. This refers to the electrical power of the air conditioner.

[0136] Indoor temperature comfort range constraints:

[0137]

[0138] in, and These represent the lower and upper limits of the comfortable indoor temperature range, respectively.

[0139] Air conditioning ramping constraints:

[0140]

[0141] in, and These represent the lower and upper limits of air conditioner power variation, respectively.

[0142] (3) Computing center constraints

[0143] IT equipment power limits and ramping constraints:

[0144]

[0145]

[0146] in, and These are the minimum and maximum power of the IT equipment, respectively. The upper limit of the gradient rate is set at 0.2 to 0.4.

[0147] Daily computing power limit:

[0148]

[0149] Minimum following ratio constraint during working hours:

[0150]

[0151] In the formula, For work hours, This represents the minimum following ratio.

[0152] Relationship between total electrical load and IT power PUE:

[0153]

[0154] in, Energy efficiency ratio for data centers; This represents the total electrical load of the computing center.

[0155] (4) Power balance constraints

[0156] Power balance in each functional area:

[0157]

[0158] in, Photovoltaic power; Wind power output; This is the power of a conventional electrical load.

[0159] Power balance of computing center:

[0160]

[0161] Shared energy storage total power balance:

[0162]

[0163] (5) Tiered carbon trading constraints

[0164] Carbon emission allowances:

[0165]

[0166] in, Total carbon emission allowance; Carbon emission allowance factor for purchasing electricity from the grid, in kg / kWh.

[0167] Actual carbon emissions:

[0168]

[0169] In the formula, For actual total carbon emissions, The actual carbon emission factor for electricity purchased from the grid is expressed in kg / kWh.

[0170] Computing power time shift reduces emissions:

[0171]

[0172] in, Reduce emissions by shifting computing power over time; For valley periods; Number of valley periods; This represents the difference in carbon emission factors between peak and off-peak periods. , The peak carbon emission factor. Carbon emission factors during off-peak periods.

[0173] Net carbon emissions:

[0174]

[0175] Introducing positive and negative carbon emission variables:

[0176]

[0177] in, For positive net carbon emissions (the portion requiring the purchase of allowances); For negative net carbon emissions (the portion of the allowance that can be sold).

[0178] Breaking down positive net carbon emissions into multiple tiers:

[0179]

[0180]

[0181]

[0182] in, The number of steps in the ladder; This represents the carbon emissions for the vth step. The interval length for each step segment, in kg.

[0183] Tiered carbon trading costs:

[0184]

[0185]

[0186] in, Total carbon trading cost; The carbon trading price for the vth tier; The base price for carbon trading; This represents the price growth rate.

[0187] (6) Soft constraint guidance

[0188] The target power of energy storage and the target power of computing power output by the upper-level Q-learning layer serve as soft constraints, and slack variables are introduced to penalize deviations:

[0189]

[0190]

[0191]

[0192] in, This is a soft constraint penalty item; For soft constraint penalty weights; and These are slack variables for energy storage and computing power, respectively. Total energy storage capacity; The target energy storage power is the output of Q-learning; The target power of computing power output by Q-learning.

[0193] (iv) Solution method

[0194] This invention employs the YALMIP toolkit to construct a mixed-integer linear programming model and calls the Gurobi solver for solution. For the upper-layer Q-learning training, an offline training method is used, iteratively training under multiple typical daily scenarios until the Q-table converges. After training, the upper-layer policy can output the target power in real time to guide the lower-layer optimization. The lower-layer optimization solves for the precise operating power of each device, achieving economical, low-carbon, and efficient system operation.

[0195] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0196] Implementation Case:

[0197] This embodiment takes a university campus in a hot-summer, cold-winter region of my country as the research object. The campus covers an area of ​​approximately 300,000 square meters, with a total building area of ​​approximately 240,000 square meters, including three main functional areas: teaching area, dormitory area, and scientific research and experimental area, with approximately 7,000 faculty and students. In terms of time and scenario parameters, four typical days are set: spring, summer, autumn, and winter. Each typical day contains 24 scheduling periods with a time resolution of 1 hour. The number of days for each typical day is weighted as follows: 91 days for spring, 91 days for summer, 91 days for autumn, and 92 days for winter. The total number of days in a year is calculated as 365 days.

[0198] The parameters for the shared energy storage power station are set as follows: charging efficiency is 0.98, discharging efficiency is 0.98, self-discharge efficiency is 0, SOC upper and lower limits are 0.9 and 0.1 respectively, initial SOC is 0.2, and energy rate is 2.8. Regarding energy storage investment costs, the unit power cost is 1000 yuan / kW, the unit capacity cost is 1200 yuan / kWh, the unit operation and maintenance cost is 72 yuan / (year·kW), and the energy storage lifespan is calculated as 10.5 years.

[0199] The computing center parameters are set as follows: maximum power of IT equipment is 500 kW, minimum power is 100 kW, ramp rate is 0.3, total daily computing power is 6000 kW·h, PUE is 1.4, latency penalty is 0.35 yuan / kW, time shift penalty is 0.12 yuan / kW, working hours are 9:00~12:00 and 14:00~17:00, and the minimum follow ratio during working hours is 0.3. The investment cost of the computing center is 800 yuan / kW, the operation and maintenance cost is 50 yuan / (year·kW), and the life cycle is 10 years.

[0200] The building parameters for each functional area are shown in Table 1.

[0201] Table 1 Building Thermal Dynamics Parameters

[0202]

[0203] The time-of-use electricity pricing is set as shown in Table 2.

[0204] Table 2 Time-of-use electricity prices

[0205]

[0206] For Q-learning training parameters, the number of SOC discretization levels is 5, the number of electricity price levels is 3, the number of new energy output levels is 3, the number of computing power demand levels is 3, the number of energy storage actions is 3, the number of computing power actions is 3, the learning rate is 0.03, the discount factor is 0.95, the initial exploration rate is 1.0, the final exploration rate is 0.01, the exploration rate decay factor is 0.98, and the number of training rounds is 500.

[0207] Regarding load and renewable energy data, outdoor temperature and solar radiation data are generated based on historical meteorological station data. The daily average temperatures for spring, summer, autumn, and winter are 16.5℃, 28.2℃, 14.3℃, and -2.1℃, respectively. Solar irradiance intensity is scaled according to seasonal scales: 0.75 times for spring, 1.00 times for summer, 0.55 times for autumn, and 0.30 times for winter.

[0208] The photovoltaic output, wind power output, and conventional electrical load data for each functional area are generated based on historical measured data. The load in the teaching area is concentrated between 8:00 and 18:00, while the load in the dormitory area is higher between 12:00 and 14:00 and after 22:00. The load in the research area is relatively evenly distributed throughout the day. The peak photovoltaic output on typical days occurs between 12:00 and 14:00 in summer, with a maximum power of about 800kW. The wind power output is higher in spring and winter, with a maximum power of about 600kW.

[0209] The computing power demand curve is generated based on the characteristics of campus scientific research computing tasks. The computing power demand is higher during working hours (9:00-12:00 and 14:00-17:00), with a peak of about 450 kW. The computing power demand is lower during nighttime hours (0:00-6:00), about 100-120 kW. The seasonal correction factors are 1.0 in spring, 0.75 in summer, 1.05 in autumn, and 0.65 in winter.

[0210] To verify the effectiveness of this invention, the following comparative simulation scheme was set up:

[0211] Table 3 Scene Settings

[0212]

[0213] Figure 1 This is a schematic diagram illustrating the principle of a smart campus shared energy storage-computing power joint optimization scheduling method considering tiered carbon trading, as per the present invention. Figure 1As shown, the system adopts a two-layer optimization architecture. The upper layer is a Q-learning strategy generation module, which selects the joint action of energy storage charging and discharging and computing power scheduling modes through the ε-greedy strategy based on the state space composed of energy storage charge state, electricity price period, new energy output level and computing power demand level, and outputs the target power of energy storage and computing power. The lower layer is a YALMIP optimization solution module, which uses the target power output by the upper layer as a soft constraint to solve for the optimal capacity configuration of shared energy storage and the operating power of each device in each period with the goal of minimizing the annualized total cost of the system.

[0214] Figure 2 The graphs show outdoor temperature curves for four typical days. As can be seen, summer days have the highest temperatures, peaking around 16:00 at approximately 33.9℃; winter days have the lowest temperatures, with the trough occurring in the early morning at approximately -9.6℃; spring and autumn days have moderate temperatures with relatively gradual changes. These differences in outdoor temperature directly affect the building's thermal dynamics and air conditioning load demand, providing a basis for subsequent analysis of scheduling strategies for different seasons.

[0215] Figure 3 , Figure 4 , Figure 5 The figures show the solar and wind load curves for four typical daytime teaching areas, research and experimental areas, and dormitory areas. Looking at the load characteristics of each functional area, the teaching area load is concentrated between 8:00 and 18:00, highly correlated with the course schedule, with peak loads occurring between 10:00 and 11:00 and between 15:00 and 16:00, reaching a maximum load of approximately 800kW. The dormitory area load exhibits a bi-peak characteristic, with higher loads during the student lunch break (12:00-14:00) and the rest period after 22:00, reaching a maximum load of approximately 650kW. The research and experimental area load is relatively evenly distributed throughout the day, with relatively gentle fluctuations, reflecting the continuous nature of scientific research and experimental work.

[0216] In terms of renewable energy output, photovoltaic (PV) power output is mainly concentrated between 7:00 and 18:00, with a peak between 12:00 and 14:00. Output is highest in summer (approximately 200-250 kW) and lowest in winter (approximately 80-100 kW). Wind power output is stronger in spring and winter, reaching a maximum of approximately 150-180 kW, while it is weaker in summer. The configuration of renewable energy varies across different functional areas. The research area has the largest PV installed capacity, followed by the dormitory area, while the teaching area has a relatively smaller capacity. This is related to the roof area and building orientation of each functional area.

[0217] Figure 6 , Figure 7 , Figure 8 The graphs show the air conditioning power and indoor temperature changes in the teaching area, research and experimental area, and dormitory area on a typical summer day.

[0218] From the results of the teaching area operation ( Figure 6From the perspective of [data missing], the air conditioning power is maintained between 120-140kW from 9:00 to 17:00, which closely matches the teaching activity period. The indoor temperature is stabilized within the comfortable range of 24-26℃ under air conditioning control, with temperature fluctuations controlled within 2℃ / h, meeting user comfort requirements. The air conditioning power ramp-up constraint is effectively met, with the power change rate between adjacent time periods not exceeding 40kW / h, avoiding the impact of frequent air conditioning start-stop cycles on equipment lifespan.

[0219] From the results of the operation of the research zone ( Figure 7 From the perspective of air conditioning power, the power consumption remained relatively stable throughout the day, at approximately 100-120kW, which is consistent with the continuous nature of scientific research experiments. Indoor temperature also remained within a comfortable range with minimal fluctuations, providing stable environmental conditions for precision experimental equipment.

[0220] Based on the operational results of the dormitory area ( Figure 8 From the perspective of data analysis, the air conditioning power exhibits a clear bi-peak characteristic, with a peak of approximately 150kW during the student lunch break (12:00-14:00) and a peak of approximately 140kW during the rest period (22:00-2:00). Indoor temperature is effectively controlled during both lunch breaks and nighttime, demonstrating the invention's precise response to the energy needs of the dormitory area.

[0221] Figure 9 , Figure 10 , Figure 11 and Figure 12 The graphs show the scheduling curves of computing centers during four typical days in spring, summer, autumn, and winter. Figure 13 This is a time-shift curve of computing load. To accurately reflect the energy source of the computing load, Figures 9 to 12 The bar chart only shows the power supply composition that meets the total load of the computing center, which is jointly supplied by grid power purchase and shared energy storage discharge; system-level energy storage replenishment behaviors such as reverse charging of shared energy storage from the computing center are not included in the computing power supply composition, and the overall charging and discharging process of shared energy storage is handled separately. Figure 14 Unified display.

[0222] from Figure 9During a typical spring day, the baseline total load exhibits a bi-peak characteristic during the day and a lower load at night. The optimized total load curve demonstrates a strict "peak avoidance" characteristic, with the optimized total load consistently less than or equal to the baseline total load throughout the day. Particularly during the evening peak electricity price period from 18:00 to 21:00, the optimized total load shows a significant voltage drop, for example, decreasing from approximately 480kW to 210kW at 19:00. The system fully utilizes the elasticity and latency allowed by the computing power task, effectively removing load during high-price periods, and does not show significant excess time-shift increases in other time periods. From the perspective of power supply composition, during the peak electricity price periods from 9:00 to 12:00 and from 18:00 to 21:00, the shared energy storage system discharged significantly to support the computing center, significantly reducing dependence on grid power purchases.

[0223] from Figure 10 During typical summer days, driven by strong photovoltaic output, the computing center exhibits a strong "load shift" characteristic. During the peak photovoltaic power generation period from 12:00 to 15:00, the optimized total load significantly exceeds the baseline load, forming a peak power consumption of approximately 700kW. This indicates that the model actively shifts a large amount of computing power tasks to this period for concentrated processing, fully utilizing low-carbon and economical new energy sources. Conversely, during the evening period from 18:00 to 21:00, when there is no peak solar power, the load is deeply reduced to a minimum of approximately 140kW, primarily satisfied by shared energy storage discharge, achieving deep collaborative arbitrage.

[0224] from Figure 11 On a typical autumn day, the scheduling pattern is similar to that of spring, with the optimized total load throughout the day strictly controlled within the baseline total load line. Computational tasks were significantly reduced and delayed during the evening peak electricity price period from 18:00 to 21:00 to avoid high electricity purchase costs and carbon emission costs. On the power supply side, the discharge power of the shared energy storage precisely covered the peak periods from 9:00 to 12:00 during the day and from 18:00 to 21:00 at night, effectively playing a role in smoothing out peak electricity purchases from the grid.

[0225] from Figure 12 On a typical winter day, due to the low outdoor temperature, the cooling load of the computing center is significantly reduced, and the overall baseline load is at its lowest level of the year. The system also implemented significant load reduction during the peak electricity price periods around 11:00 and 20:00. In terms of power supply composition, the shared energy storage accounts for a very high proportion of discharge during peak electricity price periods, almost completely meeting the energy demand of the computing center after load reduction during some high-price periods (such as 11:00 and 19:00). This resulted in the computing center's grid-purchased electricity approaching zero during these periods, achieving optimal economic and carbon-reducing operation of the system as a whole.

[0226] Figure 13The computing load time-shifting curve further illustrates the time-shifting effect. During off-peak hours (0:00-7:00), IT power increased by an average of approximately 280kW compared to the original demand, while during peak hours (9:00-11:00 and 18:00-21:00), IT power decreased by an average of approximately 180kW compared to the original demand, achieving a power transfer effect of approximately 460kW. This reduces the electricity purchase cost during peak hours and indirectly reduces emissions by shifting the load to low-carbon emission periods. The time-shifting penalty cost is approximately 38 yuan / day, and the delay penalty cost is approximately 22 yuan / day, both within a reasonable range, indicating that the scheduling strategy effectively balances economic efficiency with the quality of computing power services.

[0227] Figure 14 The graph shows the power consumption on a typical day on campus and the charging and discharging power of shared energy storage, illustrating the synergistic effect of energy storage dispatch on typical summer days.

[0228] As shown in the figure, the peak total power purchase on campus occurs between 10:00-11:00 and 19:00-20:00, reaching approximately 3200kW. Shared energy storage charges at approximately 500kW during the off-peak hours (0:00-7:00) and discharges at approximately 400-600kW during the peak hours (9:00-11:00 and 18:00-21:00). The energy storage discharge period covers 60%-70% of the peak power purchase period, effectively reducing the peak power purchase by approximately 420kW, a reduction of 13.1%. The power traded between energy storage and different functional areas shows significant differences. The teaching and research areas mainly purchase electricity from energy storage during the day, while the dormitory area mainly purchases electricity from energy storage in the evening, reflecting the differences in load characteristics between different functional areas and verifying the guiding role of thermal dynamic constraints on energy storage dispatch.

[0229] The energy storage operation strategy fully reflects the scheduling principle of "charging during off-peak hours and discharging during peak hours," which is consistent with the incentive direction of time-of-use pricing. The energy storage discharge period closely coincides with the peak load period on campus, achieving peak shaving and valley filling effects and effectively reducing electricity purchase costs.

[0230] Figure 15 The training convergence curve for Q-learning shows the cumulative reward changes for each typical day over 500 training rounds.

[0231] As shown in the graph, the cumulative reward increases rapidly in the initial stage, enters a stable period after about 150 rounds, and finally stabilizes between -85 and -75. The convergence speed is faster in spring and autumn, reaching a stable period after about 120 rounds; the convergence speed is relatively slower in summer due to higher load and greater fluctuations in new energy sources, stabilizing after about 180 rounds; the convergence speed in winter is between that of spring / autumn and summer. The exploration rate gradually decreases from an initial 1.0 to 0.01, achieving a smooth transition from full exploration to utilization.

[0232] After training, the optimal actions for each state were extracted, and analysis revealed that during peak electricity prices and with high State of Charge (SOC), energy storage tends to discharge, while computing power tends to delay (reducing IT power). During off-peak electricity prices and with low SOC, energy storage tends to charge, while computing power tends to operate at full capacity (increasing IT power). This strategy embodies the collaborative scheduling concept of "charging energy storage during off-peak hours and operating computing power at full capacity, while discharging and supplying electricity during peak hours and delaying computing power," which aligns with the incentive direction of time-of-use pricing and carbon trading prices, thus validating the rationality of the Q-learning strategy.

[0233] The annualized total cost and carbon emissions of each option are compared in Table 4.

[0234] Table 4 Comparison of total cost and carbon emissions for each option

[0235]

[0236] The comparison results in Table 4 show that:

[0237] Option 2 (energy storage only) reduces annual costs by 7.07% and emissions by 8.14% compared to Option 1. Shared energy storage effectively reduces electricity purchase costs through peak shaving and valley filling, while also reducing carbon emissions from grid-based electricity purchases.

[0238] Option 3 (computing power only) reduces annual costs by 3.39% and emissions by 4.26% compared to Option 1. Although computing power time-shifting also generates some benefits, the effect is weaker than energy storage scheduling, indicating that relying solely on computing power flexibility is insufficient to fully realize the system's regulation potential.

[0239] Option 4 (energy storage + computing power) reduces annual costs by 11.33% and emissions by 14.93% compared to Option 1. The benefits of joint scheduling of energy storage and computing power are greater than the sum of their individual benefits, verifying the necessity of coordinated scheduling.

[0240] Option 5 (the present invention) reduces annual costs by 14.08% and emissions by 20.12% compared to Option 1; and further reduces annual costs by 3.09% and emissions by 6.12% compared to Option 4. The introduction of the tiered carbon trading mechanism improves both environmental benefits and economic efficiency, achieving a win-win situation for both environmental protection and the economy.

[0241] This embodiment uses a university campus as an example to construct a smart campus shared energy storage-computing power joint optimization scheduling system that considers tiered carbon trading. Simulation results verify the effectiveness and superiority of the invention. The results show that:

[0242] (1) The upper-level Q-learning strategy can learn the optimal energy storage and computing power joint scheduling mode and generate an effective target power guidance signal;

[0243] (2) The optimized configuration of shared energy storage can realize "valley storage and peak release", effectively reducing the cost of electricity purchase, and can generate a combined benefit of 1+1>2 when coordinated with computing power scheduling;

[0244] (3) The time-shift scheduling of the computing center makes full use of load flexibility and achieves load transfer and indirect emission reduction while ensuring service quality;

[0245] (4) The tiered carbon trading mechanism effectively incentivizes the system to control carbon emissions, improving both environmental benefits and economic efficiency.

[0246] (5) Compared with the benchmark solution, the annual comprehensive cost of the present invention is reduced by 14.08% and the carbon emissions are reduced by 20.12%, which verifies the effectiveness and advancement of the technical solution.

[0247] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0248] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A smart campus shared energy storage-computing power joint optimization scheduling method considering tiered carbon trading, characterized in that: include, Obtain operating parameters of the campus integrated energy system, building thermal dynamic parameters, computing center parameters, and tiered carbon trading parameters; A two-layer optimization scheduling system is constructed, consisting of an upper-layer policy generation module and a lower-layer optimization solution module. In the upper-layer policy generation module, the state space and action space are constructed based on the Q-learning algorithm, the joint action is selected through the ε-greedy policy and multiple rounds of training are performed, and the target power of energy storage and the target power of computing power are output based on the acquired parameters. In the lower-level optimization solution module, a mixed-integer linear programming model is constructed based on YALMIP. Guided by the target power output by the upper-level strategy generation module, and with the goal of minimizing the annualized total cost of the system, the optimal capacity configuration of shared energy storage and the operating power of each device in each time period are solved. Output and display the energy storage configuration results, the air conditioning operation and temperature changes in each functional area, the computing center scheduling curve, and the carbon trading allocation results.

2. The method as described in claim 1, characterized in that: In the upper-layer strategy generation module, the state space includes four dimensions: energy storage charge state level, electricity price time period type, total output level of new energy, and benchmark load level of computing center; the action space is composed of energy storage actions and computing power actions. Energy storage actions include three modes: charging, discharging, and holding, while computing power actions include three modes: full load, low load, and delay.

3. The method as described in claim 1, characterized in that: In the upper-level strategy generation module, an ε-greedy strategy is used for action selection. The exploration rate decreases as the number of training rounds increases, and an immediate reward is calculated after each action is executed in each time period. The immediate reward includes electricity purchase cost, carbon trading cost, computing power delay penalty, computing power time shift penalty, and SOC out-of-bounds penalty.

4. The method as described in claim 1, characterized in that: In the lower-level optimization solution module, the annualized total system cost includes the investment and operation and maintenance costs of shared energy storage, the investment and operation and maintenance costs of computing power center, the cost of time-of-use electricity purchase, the transaction costs of shared energy storage with each functional area and computing power center, the cost of computing power delay and time shift penalty, and the cost of tiered carbon trading.

5. The method as described in claim 1, characterized in that: The lower-level optimization solution module includes constraints such as the shared energy storage charge state evolution and rate constraint, building thermal dynamic equation constraint for each functional area, indoor temperature comfort range constraint, air conditioning power ramping constraint, upper and lower limits and ramping constraint of IT equipment power in the computing center, daily computing power conservation constraint, minimum following ratio constraint during working hours, power balance constraint, and tiered carbon trading constraint.

6. The method as described in claim 5, characterized in that: The building thermal dynamic equation constraints are constructed based on an equivalent thermal parameter model, including the thermal balance relationship between wall heat capacity, wall thermal resistance, solar radiation, indoor and outdoor temperature difference, and air conditioning heating / cooling power.

7. The method as described in claim 1, characterized in that: The tiered carbon trading parameters include the carbon trading base price, price growth rate, segment interval length, grid purchase carbon emission quota factor, and actual carbon emission factor. The lower-level optimization solution module decomposes net carbon emissions into multiple tiers by introducing positive and negative carbon emission variables and segment variables, and calculates carbon trading costs using incremental carbon trading prices.

8. The method as described in claim 1, characterized in that, The lower-level optimization solution module is also used to quantify the emission reduction benefits of computing power time shift. By calculating the difference in the grid carbon emission factor during peak and valley periods, the carbon emission reduction caused by the shift of computing power load from peak to valley periods is included in the carbon emission reduction.

9. The method as described in claim 1, characterized in that: The target power of energy storage and the target power of computing power output by the upper-layer strategy generation module serve as soft constraints for guidance, and the deviation is penalized by introducing slack variables.

Images