Intelligent scheduling method and system for multi-energy cooperative heat supply

Abstract

The invention relates to the technical field of heat supply, and provides an intelligent scheduling method and system for multi-energy cooperative heat supply, and the method comprises the steps: constructing a multi-time scale division mechanism, processing time scales, and dividing the time scales into a long-term time scale, a medium-term time scale and a short-term time scale; aiming at a long-term time scale, constructing an LSTM-Transfer hybrid model to predict renewable energy output and thermal load demand; for the mid-term time scale, dynamic rolling optimization is adopted, and multi-objective optimization is achieved through an NSGA-III algorithm; and for the short-term time scale, constructing a deep reinforcement learning DRL model, and performing dynamic compensation control through the deep reinforcement learning DRL model. According to the method, different processing strategies are used for different time scales, collaborative optimization can be carried out on multi-energy heat supply on multiple time scales, and the overall scheduling effect is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of heating, and in particular, to an intelligent scheduling method and system for multi-energy collaborative heating. Background Art

[0002] Currently, most heating systems adopt simple combination modes such as solar energy + gas boiler or electric energy + ground source heat pump, lacking dynamic collaborative scheduling of diversified energy sources such as waste heat recovery and biomass energy, resulting in insufficient utilization of energy complementarity.

[0003] Application No. CN202210877516.9 discloses a multi-heat-source networked heating collaborative optimization scheduling system based on blockchain, which achieves a good fit with the multi-heat-source networked heating collaborative scheduling through blockchain technology, providing safe and efficient services for multi-heat-source networked scheduling. Another example is Application No. CN202211711939.X, which discloses a method for collaborative optimization scheduling of boiler-turbine-network in cogeneration district heating, which assigns collaborative optimization scheduling parameters to the corresponding control parameters of boiler APC, turbine DEH, and desuperheater pressure reducer, realizing automatic judgment, adjustment, and control of the system, and quickly responding to the load changes of heat network users. Still another example is Application No. CN202210864193.X, which discloses a method for collaborative optimization scheduling of multi-heat-source heating system based on hierarchical reinforcement learning. It uses the value function obtained by reinforcement learning to evaluate actions, combines the behavior sequences between each heat source, selects the optimal strategy, realizes the collaborative optimization scheduling of the multi-heat-source networked heating system, and improves the learning speed and efficiency.

[0004] The above several existing technologies all involve multi-energy heating collaborative optimization methods. However, they fail to solve the problem of collaborative optimization on multiple time scales, and their scheduling effects are relatively high and need to be improved. Summary of the Invention

[0005] Based on this, in order to solve the problem that the existing multi-energy heating collaborative optimization methods fail to solve the problem of collaborative optimization on multiple time scales, the present invention provides an intelligent scheduling method and system for multi-energy collaborative heating. By dividing the time scale into long-term time scale, medium-term time scale, and short-term time scale, and using different processing strategies for different time scales, it can collaboratively optimize multi-energy heating on multiple time scales and improve the overall scheduling effect. The specific technical solutions are as follows:

[0006] An intelligent scheduling method for multi-energy collaborative heating, which includes the following steps:

[0007] Construct a multi-time scale division mechanism, process the time scale, and divide it into long-term time scale, medium-term time scale, and short-term time scale;

[0008] For the long-term time scale, a hybrid LSTM-Transformer model is constructed to predict the output of renewable energy and the demand for heat load;

[0009] For the medium-term time scale, dynamic rolling optimization is adopted, and multi-objective optimization is realized through the NSGA-Ⅲ algorithm;

[0010] For the short-term time scale, a deep reinforcement learning DRL model is constructed, and dynamic compensation control is carried out through the deep reinforcement learning DRL model.

[0011] The intelligent scheduling method for multi-energy collaborative heating processes the time scale by constructing a multi-time scale division mechanism, which is divided into long-term, medium-term, and short-term time scales. For the long-term time scale, a hybrid LSTM-Transformer model is constructed to predict the output of renewable energy and the demand for heat load, which can combine spatio-temporal feature coupling and attention enhancement mechanisms. Through the complementary advantages of LSTM-Transformer, it can solve the problems of local mode loss and weakening of global correlation in the long cycle of traditional single models. For the medium-term time scale, dynamic rolling optimization is adopted, and multi-objective optimization is realized through the NSGA-Ⅲ algorithm. Using the characteristics of the NSGA-Ⅲ algorithm and adopting a rolling optimization strategy, it can not only achieve the global optimum of the economy, stability, and comfort of the heating system, avoid the overall imbalance caused by excessive optimization of a single objective, but also recalculate the optimal solution in each rolling window to respond to the fluctuations of renewable energy output and load changes in real time, improving the system robustness. For the short-term time scale, a deep reinforcement learning DRL model is constructed, and dynamic compensation control is carried out through the deep reinforcement learning DRL model. Through the autonomous decision-making ability of DRL, it can achieve precise collaborative control of the multi-energy heating system in the short time domain, providing a reliable execution basis for long / medium-term optimization.

[0012] Preferably, the specific method for constructing a hybrid LSTM-Transformer model to predict the output of renewable energy and the demand for heat load includes the following steps:

[0013] Obtain the time window feature matrix of the input sequence, and extract local temporal features through a bidirectional LSTM network;

[0014] Obtain the meteorological feature matrix, and dynamically fuse the meteorological influence through a cross-attention layer according to the obtained meteorological feature matrix;

[0015] Capture the cross-time step correlation through Transformer;

[0016] Obtain the predicted values of the output of renewable energy and the demand for heat load according to the fused meteorological influence and the captured correlation of time steps.

[0017] Preferably, dynamic rolling optimization is adopted, and the specific method for multi-objective optimization implemented by the NSGA-Ⅲ algorithm includes the following steps:

[0018] Obtain the energy cost C energy , the system loss C loss and the user temperature deviation ΔT user , and construct an objective optimization function based on the energy cost, system loss, and user temperature deviation

[0019] min{C energy , C loss , ΔT user};

[0020] Obtain the total supply of energy equipment ∑E i and the total heat load demand Q of the system load , and obtain the constraint condition s.t ∑E i ≥Q load ;

[0021] where E i represents the output energy of the i-th type of energy equipment.

[0022] Preferably, the specific method for constructing a deep reinforcement learning DRL model and performing dynamic compensation control through the deep reinforcement learning DRL model includes the following steps:

[0023] Obtain the state variables and action variables of the heat pump system, and obtain the reward function according to the state variables and action variables;

[0024] Obtain the discount factor, and obtain the optimal policy according to the discount factor and the reward function;

[0025] Perform dynamic compensation control of the system according to the optimal policy.

[0026] Preferably, the intelligent scheduling method for multi-energy collaborative heating further includes the following steps:

[0027] Construct an energy priority matrix based on game theory;

[0028] According to the constructed energy priority matrix based on game theory, match different scenarios through transfer learning.

[0029] An intelligent scheduling system for multi-energy collaborative heating, which is used to implement the intelligent scheduling method for multi-energy collaborative heating, includes a time scale division module. The time scale division module is used to construct a multi-time scale division mechanism, process the time scale, and divide it into a long-term time scale, a medium-term time scale, and a short-term time scale. The time scale division module includes:

[0030] Long-term time scale unit, used to construct a prediction of renewable energy output and heat load demand based on an LSTM-Transformer hybrid model for the long-term time scale;

[0031] Medium-term time scale unit, used to perform dynamic rolling optimization for the medium-term time scale and achieve multi-objective optimization through the NSGA-Ⅲ algorithm;

[0032] Short-term time scale unit, used to construct a deep reinforcement learning DRL model for the short-term time scale and perform dynamic compensation control through the deep reinforcement learning DRL model.

[0033] Preferably, the long-term time scale unit includes:

[0034] Local time series feature acquisition sub-unit, used to obtain the time window feature matrix of the input sequence and extract local time series features through a bidirectional LSTM network;

[0035] Meteorological impact acquisition sub-unit, used to obtain the meteorological feature matrix and dynamically fuse the meteorological impact through a cross-attention layer according to the obtained meteorological feature matrix;

[0036] Relevance acquisition sub-unit, used to capture the relevance across time steps through Transformer;

[0037] Prediction sub-unit, used to obtain the predicted values of renewable energy output and heat load demand according to the fused meteorological impact and the captured relevance of time steps.

[0038] Preferably, the medium-term time scale unit includes:

[0039] Objective optimization function acquisition sub-unit, used to obtain the energy cost C energy 、system loss C loss and user temperature deviation ΔT user , and construct the objective optimization function min{C energy ,C loss ,ΔT user} according to the energy cost, system loss and user temperature deviation;

[0040] Constraint condition acquisition sub-unit, used to obtain the total energy equipment supply ∑E i and the total system heat load demand Q load , and obtain the constraint condition s.t∑E i ≥Q load ;

[0041] where E i represents the output energy of the i-th type of energy equipment.

[0042] Preferably, the short - term time - scale unit includes:

[0043] A reward function acquisition subunit, configured to acquire state variables and action variables of the heat pump system, and obtain a reward function according to the state variables and the action variables;

[0044] An optimal policy acquisition subunit, configured to acquire a discount factor, and obtain an optimal policy according to the discount factor and the reward function;

[0045] A dynamic compensation control subunit, configured to perform dynamic compensation control of the system according to the optimal policy.

[0046] Preferably, the intelligent scheduling system for multi - energy collaborative heating further includes:

[0047] An energy priority matrix construction subunit, configured to construct an energy priority matrix based on game theory;

[0048] A scenario matching subunit, configured to match different scenarios through transfer learning according to the constructed energy priority matrix based on game theory. Description of the Drawings

[0049] The present invention can be further understood from the following description in conjunction with the drawings. The components in the drawings are not necessarily drawn to scale, but the emphasis is placed on showing the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.

[0050] Figure 1 is the overall process schematic diagram of an intelligent scheduling method for multi - energy collaborative heating in an embodiment of the present invention;

[0051] Figure 2 is the process schematic diagram of a specific method for predicting the output of renewable energy and heat load demand in an embodiment of the present invention;

[0052] Figure 3 is the process schematic diagram of a specific method for realizing multi - objective optimization through the NSGA - Ⅲ algorithm in an embodiment of the present invention;

[0053] Figure 4 is the process schematic diagram of a specific method for performing dynamic compensation control through a deep reinforcement learning DRL model in an embodiment of the present invention;

[0054] Figure 5 is the process schematic diagram of an intelligent scheduling method for multi - energy collaborative heating in another embodiment of the present invention;

[0055] Figure 6 is the overall structure schematic diagram of an intelligent scheduling system for multi - energy collaborative heating in an embodiment of the present invention. Detailed Embodiments

[0056] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with its embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not limit the protection scope of the present invention.

[0057] It should be noted that when an element is referred to as being "fixed to" another element, it can be directly on the other element or there can also be an intermediate element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be an intermediate element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are only for the purpose of illustration and do not represent the only implementation.

[0058] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs. The terms used in the description of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0059] The "first" and "second" mentioned in the present invention do not represent specific quantities and sequences, but are only used for name distinction.

[0060] As Figure 1 shown, an embodiment of the present invention provides an intelligent scheduling method for multi-energy collaborative heating, which includes the following steps:

[0061] S1. Construct a multi-time scale division mechanism, process the time scale, and divide it into a long-term time scale, a medium-term time scale, and a short-term time scale.

[0062] For the specific division of the long-term time scale, medium-term time scale, and short-term time scale, the long-term time scale is greater than the medium-term time scale, and the medium-term time scale is greater than the short-term time scale. Specifically, the long-term time scale can be set to 24 hours to 72 hours. Correspondingly, the medium-term time scale is set to 1 hour to 24 hours, and the short-term time scale is set to 1 minute to 1 hour. Of course, according to the actual situation, the scale ranges of the long-term time scale, medium-term time scale, and short-term time scale can be adjusted, which will not be elaborated here.

[0063] The multi-time scale division mechanism can be understood as a preset multi-time scale division strategy, which can dynamically adjust the range of the divided time scales according to factors such as the meteorological environment of the target heating area of the intelligent scheduling system or the target heating group (i.e., the audience group within the heating area). For example, dynamically adjust the ranges of different time scales according to the order of magnitude of the audience or the control accuracy required by the system.

[0064] S2. For the long-term time scale, construct a prediction model for renewable energy output and heat load demand based on the LSTM-Transfermer hybrid model.

[0065] Preferably, in step S2, as Figure 2 shown, the specific method for constructing a prediction model for renewable energy output and heat load demand based on the LSTM-Transformer hybrid model includes the following steps:

[0066] S21. Obtain the time window feature matrix X of the input sequence t-24:t , and extract the local time series feature LSTM(X t-24:t ) through a bidirectional LSTM network; among them, assuming that the time resolution is 1 hour, the time window feature matrix represents the historical data from the past 24 hours to the current moment t, and its data structure is represented as R N×T×D , N represents the batch size or the number of samples, T = 24 represents the time step, which is a 24-hour historical window, and D represents the feature dimension, including dynamic parameters such as temperature, historical load, and equipment status. LSTM(X t-24:t ) filters out noise interference through a gating mechanism, controls the memory retention rate through a forget gate, and updates the cell state through an input gate.

[0067] S22. Obtain the meteorological feature matrix W weather = [T future × D weather , and dynamically fuse the meteorological influence Attention(W weather ) according to the obtained meteorological feature matrix.

[0068] Specifically, the meteorological feature matrix contains variables such as temperature, humidity, wind speed, and irradiance, and the dimension is W weather = [T future × D weather , usually 4-6 dimensions, and the dimension difference can be eliminated by sliding window normalization.

[0069] S23. Capture the cross-time step correlation through Transformer.

[0070] Specifically, for the h output by LSTM tPerform global feature enhancement: capture the correlations across time steps through the multi-head attention mechanism (such as the lagging impact of extreme weather on load), and output the encoded vector E t = Transformer(h t )

[0071] S24. According to the fused meteorological impact and the captured correlations of time steps, obtain the predicted values of renewable energy output and heat load demand

[0072] Specifically, the predicted value where the predicted value represents the combined predicted value of renewable energy output (such as wind power and photovoltaic) and heat load demand at time point t, with the unit of kW or MW.

[0073] By the fused meteorological impact and the captured correlations of time steps, obtain the predicted values of renewable energy output and heat load demand. It jointly processes the time window feature matrix and the meteorological feature matrix, captures the short-term dynamics of the device through LSTM and models the long-term correlations of the region through Transformer, realizing multi-scale feature fusion and overcoming the limitations of a single data source. In addition, the LSTM-Transformer hybrid structure takes into account both local temporal dependencies (LSTM) and global pattern capture (Transformer), reducing the prediction error by 18%-25% compared with a single model.

[0074] Capture the correlations across time steps through the multi-head attention mechanism (such as the lagging impact of extreme weather on load), which can automatically identify key meteorological factors and improve the prediction robustness under extreme weather.

[0075] S3. For the medium-term time scale, adopt dynamic rolling optimization and achieve multi-objective optimization through the NSGA-Ⅲ algorithm.

[0076] Preferably, in step S3, as Figure 3 shown, the specific method of adopting dynamic rolling optimization and achieving multi-objective optimization through the NSGA-Ⅲ algorithm includes the following steps:

[0077] S31. Obtain the energy cost C energy , system loss C loss and user temperature deviation ΔT user , and construct the objective optimization function min{C energy , C loss , ΔT user} according to the energy cost, system loss and user temperature deviation.

[0078] The energy cost is the comprehensive energy procurement cost for system operation, which includes the weighted calculation of the real-time prices of energy such as electricity, natural gas, and renewable energy. System losses represent the energy losses during the energy conversion and transmission and distribution processes, including the efficiency decay of heat pumps, pipeline heat losses, and self-discharge of energy storage. The user temperature deviation represents the root mean square deviation between the actual heating temperature of the user and the set temperature, reflecting the quality of the heating service.

[0079] For the energy cost, system losses, and user temperature deviation, standardization processing can be carried out to eliminate the influence of dimensional differences on the optimization results.

[0080] S32. Obtain the total supply of energy equipment ∑E i and the total heat load demand Q of the system load , and obtain the constraint condition s.t ∑E according to the total supply of energy equipment and the total heat load demand of the system i ≥Q load ;

[0081] where E i represents the output energy of the i-th type of energy equipment, and the total heat load demand of the system represents the total heat load demand of the system, which can be calculated in real time by the load prediction model. Preferably, in the rolling optimization, the total heat load demand of the system is dynamically adjusted with the prediction update, and the ramp rate constraint of E i needs to be satisfied.

[0082] That is to say, for the medium-term time scale, adopting dynamic rolling optimization and realizing multi-objective optimization through the NSGA-Ⅲ algorithm can be summarized as the following function:

[0083]

[0084] Here, by using the characteristics of the NSGA-Ⅲ (Nondominated Sorting Genetic Algorithm III) algorithm and adopting the rolling optimization strategy, not only can the global optimum of the economy, stability, and comfort of the heating system be realized, avoiding the overall imbalance caused by excessive optimization of a single objective, but also the optimal solution can be recalculated within each rolling window to respond to the fluctuations of renewable energy output and load changes in real time, improving the robustness of the system.

[0085] S4. For the short-term time scale, construct a deep reinforcement learning DRL model and perform dynamic compensation control through the deep reinforcement learning DRL model.

[0086] Preferably, as Figure 4As shown, in step S4, a deep reinforcement learning (DRL) model is constructed, and the specific method for dynamic compensation control through the DRL (Deep Reinforcement Learning) model includes the following steps:

[0087] S41. Obtain the state variable s of the heat pump system t and the action variable a t , and obtain the reward function r(s t , a t ) according to the state variable and the action variable.

[0088] For the state variable, it represents the operating state of the heat pump system at time t, including: environmental parameters (outdoor temperature, wind speed, etc.), equipment status (compressor frequency, water pump flow rate, etc.), and load demand (real-time heat load, temperature deviation at the user end).

[0089] S42. Obtain the discount factor γ, and obtain the optimal policy π according to the discount factor and the reward function * (a|s) = argmax π E[∑γ t r(s t , a t )].[[]END]]

[0090] The value range of the discount factor is [0.9, 0.99][0.9, 0.99], which reflects the attenuation degree of future rewards. In the heating system, the thermal inertia characteristics need to be considered: a high γ value is suitable for the heat storage system, and a low γ value is used for fast response scenarios.

[0091] S43. Perform dynamic compensation control of the system according to the optimal policy.

[0092] Among them, π(a|s) represents the probability distribution of selecting action a in state s, π * (a|s) represents the optimal policy, and argmax π means to find the policy π that maximizes the expected return, and E represents the expectation operator.

[0093] Here, through the autonomous decision-making ability of DRL, the precise coordinated control of the multi-energy heating system in a short time domain can be realized, providing a reliable execution basis for long / medium-term optimization. This DRL framework successfully solves the problem of insufficient adaptability of traditional control methods in fast time-varying and multi-disturbance scenarios, and forms a complete closed loop with long-term time-scale prediction and medium-term time-scale optimization.

[0094] After obtaining the function formulas for the long-term time scale, medium-term time scale, and short-term time scale, the heating system is controlled according to the function formulas corresponding to the divided time scales, so as to realize the intelligent scheduling of multi-energy collaborative heating.

[0095] The intelligent scheduling method for multi - energy collaborative heating processes the time scale by constructing a multi - time - scale partitioning mechanism, dividing it into long - term, medium - term, and short - term time scales. For the long - term time scale, a hybrid LSTM - Transformer model is constructed to predict the output of renewable energy and the heat load demand. It can combine spatio - temporal feature coupling and attention enhancement mechanisms, and through the complementary advantages of LSTM - Transformer, solve the problems of local pattern loss and global correlation weakening in the long - cycle of traditional single models. For the medium - term time scale, dynamic rolling optimization is adopted, and multi - objective optimization is achieved through the NSGA - Ⅲ algorithm. Utilizing the characteristics of the NSGA - Ⅲ algorithm and adopting a rolling optimization strategy, it can not only achieve the global optimum of the economy, stability, and comfort of the heating system, avoid the overall imbalance caused by over - optimization of a single objective, but also recalculate the optimal solution in each rolling window to respond to the fluctuations in the output of renewable energy and load changes in real - time, improving the system's robustness. For the short - term time scale, a deep reinforcement learning (DRL) model is constructed, and dynamic compensation control is carried out through the DRL model. Through the autonomous decision - making ability of DRL, it can achieve precise collaborative control of the multi - energy heating system in the short time domain, providing a reliable execution basis for long / medium - term optimization.

[0096] As a preferred technical solution, as Figure 5 shown, the intelligent scheduling method for multi - energy collaborative heating further includes the following steps:

[0097] S5, construct an energy priority matrix based on game theory;

[0098] S6, according to the constructed energy priority matrix based on game theory, match different scenarios through transfer learning.

[0099] Specifically, the energy priority matrix based on game theory obtains the scoring function of energy equipment Adjust the weight coefficients in the scoring function based on the scoring parameters to match different scenarios.

[0100] Among them, P avail represents the actual available power of the i - th type of energy equipment in the current time period, which can be obtained by collecting the output of renewable energy through a real - time monitoring system. P max represents the upper limit of the installed capacity of the i - th type of energy equipment, which is determined by the rated power of the equipment and the system topology structure, such as the maximum heating capacity of a ground - source heat pump unit. C unit represents the current unit energy cost (yuan / kWh or yuan / GJ) of the i - th type of energy equipment. C baseDenote the preset benchmark energy cost (yuan / kWh or yuan / GJ), which usually takes the regional average energy cost or the policy guidance price, and is processed by normalization for calculation. σ predict Denote the standard deviation of the output prediction fluctuation of the i-th type of energy equipment, which can be calculated based on the statistics of historical prediction errors (such as based on the output prediction standard deviation of a Bayesian neural network), reflecting the stability of energy supply. For example, for photovoltaic prediction σ predict Can reach 20%-30% of the installed capacity, for geothermal energy σ predict Is usually less than 5%.

[0101] α, β, and λ denote weight coefficients, which are used to dynamically adjust the energy priority and can be adjusted through transfer learning to match different scenarios. For the cold start stage of the system, α = 0.5, β = 0.3, λ = 0.2. During the operation stage, it can be dynamically adjusted based on the LSTM network to adapt to scenarios such as sudden weather changes and load surges.

[0102] Assume that the energy equipment types include two types: photovoltaic and gas boiler, and the values of α, β, and λ are set to 0.6, 0.25, and 0.15 respectively.

[0103] For photovoltaic: Score(E1) = 0.6×0.8 + 0.25×0.6 + 0.15×0.15 = 0.63.

[0104] For gas boiler: Score(E2) = 0.6×1.0 + 0.25×1.2 + 0.15×0..02 = 0.90.

[0105] At this time, the system will preferentially call the gas boiler to meet the base load, and the photovoltaic will be used as a supplementary energy source. Through the said scoring function Quantitatively evaluate the availability, economy, and stability of each energy source, which can provide a decision-making basis for the scheduling of multi-energy collaborative heating.

[0106] An embodiment of the present invention also provides an intelligent scheduling system for multi-energy collaborative heating, which is used to implement the intelligent scheduling method of multi-energy collaborative heating, as Figure 6 Shown, which includes a time scale division module. The time scale division module is used to construct a multi-time scale division mechanism, process the time scale, and divide it into a long-term time scale, a medium-term time scale, and a short-term time scale. The time scale division module includes a long-term time scale unit, a medium-term time scale unit, and a short-term time scale unit.

[0107] The long-term time scale unit is used to construct a prediction of renewable energy output and heat load demand based on the LSTM-Transfermer hybrid model for the long-term time scale.

[0108] Preferably, the long-term time-scale unit includes a local time-series feature acquisition subunit, a meteorological impact acquisition subunit, a correlation acquisition subunit, and a prediction subunit.

[0109] The local time-series feature acquisition subunit is used to obtain the time window feature matrix of the input sequence and extract local time-series features through a bidirectional LSTM network; the meteorological impact acquisition subunit is used to obtain the meteorological feature matrix, and according to the obtained meteorological feature matrix, dynamically fuse the meteorological impact through a cross-attention layer.

[0110] The correlation acquisition subunit is used to capture the correlation across time steps through a Transformer; the prediction subunit is used to obtain the predicted values of renewable energy output and heat load demand according to the fused meteorological impact and the captured correlation across time steps.

[0111] Specifically, the predicted value Among them, the predicted value represents the joint predicted value of renewable energy output (such as wind power, photovoltaic) and heat load demand at time point t, and the unit is kW or MW.

[0112] LSTM (Long Short-Term Memory network) is a special recurrent neural network that can effectively handle long-term dependence problems in sequence data. It controls the flow of information through a gating mechanism (input gate, forget gate, and output gate), and performs well in processing time series data. For example, when processing data such as renewable energy output and heat load demand that change over time, it can capture the long-term trends and periodic changes in the data.

[0113] The Transformer is a model based on the attention mechanism. It can process sequence data in parallel, avoiding the problems of gradient disappearance and low computational efficiency of traditional recurrent neural networks when processing long sequences. The attention mechanism allows the model to automatically focus on other parts of the sequence related to the current time step when processing data at each time step, thereby better capturing the global information in the data.

[0114] Combining LSTM and Transformer can give full play to the advantages of both. LSTM is responsible for capturing the local time dependencies of the data, while the Transformer is used to capture the global dependencies of the data, thereby improving the prediction accuracy of renewable energy output and heat load demand.

[0115] For the acquisition of the predicted value of renewable energy output (such as wind power, photovoltaic), it generally includes the following three steps:

[0116] Data preprocessing: Collect historical output data of renewable energy (such as solar energy, wind energy, etc.), including relevant factors such as power, wind speed, and light intensity. Clean the data, remove outliers and missing values, and perform normalization to map the data to the [0, 1] interval to improve the training efficiency and stability of the model.

[0117] Model training: Input the preprocessed data into the LSTM-Transformer model for training. During the training process, use historical data as input and the corresponding actual output values as output, and continuously adjust the model's parameters through the backpropagation algorithm to minimize the error between the predicted output and the actual output of the model.

[0118] Prediction process: After the model training is completed, input the relevant input data for a future period of time (such as wind speed, light intensity, etc. in weather forecasts) into the model, and the model can output the predicted output value of renewable energy.

[0119] For obtaining the predicted value of the heat load demand, it generally includes the following three steps:

[0120] Data collection and processing: Collect historical data of heat load demand, including relevant factors such as temperature, humidity, date, and time. Similarly, perform preprocessing operations such as data cleaning and normalization on the data.

[0121] Model adaptation: Since the characteristics of heat load demand data may be different from those of renewable energy output data, it is necessary to make appropriate adjustments to the LSTM-Transformer model. For example, hyperparameters such as the number of model layers and neurons can be adjusted to better adapt to the characteristics of heat load demand data.

[0122] Prediction and evaluation: Use the trained model to predict the future heat load demand, and evaluate the prediction accuracy of the model by comparing it with the actual data. Metrics such as mean squared error (MSE) and mean absolute error (MAE) can be used to measure the performance of the model.

[0123] By fusing the meteorological impacts and capturing the correlations of time steps, the predicted values of renewable energy output and heat load demand are obtained. It jointly processes the time window feature matrix and the meteorological feature matrix, captures the short-term dynamics of the device through LSTM and models the long-term associations of the region through Transformer to achieve multi-scale feature fusion, overcoming the limitations of a single data source. In addition, the LSTM-Transformer hybrid structure takes into account both local temporal dependencies (LSTM) and global pattern capture (Transformer), and the prediction error is reduced by 18%-25% compared with a single model.

[0124] The medium - term time - scale unit is used for the medium - term time - scale, adopting dynamic rolling optimization and realizing multi - objective optimization through the NSGA - Ⅲ algorithm.

[0125] Preferably, the medium - term time - scale unit includes: an objective optimization function acquisition subunit, which is used to acquire the energy cost C energy , the system loss C loss and the user temperature deviation ΔT user , and constructs an objective optimization function min{C energy , C loss , ΔT user} according to the energy cost, the system loss and the user temperature deviation; and a constraint condition acquisition subunit, which is used to acquire the total supply of energy equipment ∑E i and the total heat load demand Q of the system load , and obtains a constraint condition s.t ∑E i ≥Q load .

[0126] Among them, E i represents the output energy of the i - th type of energy equipment.

[0127] That is to say, for the medium - term time - scale, adopting dynamic rolling optimization and realizing multi - objective optimization through the NSGA - Ⅲ algorithm can be summarized as the following function:

[0128]

[0129] Here, by using the characteristics of the NSGA - Ⅲ algorithm and adopting a rolling optimization strategy, not only can the global optimum of the economy, stability and comfort of the heating system be achieved, avoiding the overall imbalance caused by excessive optimization of a single objective, but also the optimal solution can be recalculated within each rolling window to respond in real - time to the fluctuations of renewable energy output and load changes, improving the system robustness.

[0130] The short - term time - scale unit is used for the short - term time - scale, constructing a deep reinforcement learning (DRL) model and performing dynamic compensation control through the deep reinforcement learning (DRL) model.

[0131] Preferably, the short - term time - scale unit includes: a reward function acquisition subunit, which is used to acquire the state variable s of the heat pump system t and the action variable a t , and obtains a reward function r(s t , a t ) according to the state variable and the action variable; an optimal policy acquisition subunit, which is used to acquire the discount factor γ, and obtains the optimal policy π * (a|s)=argmax π E[∑γt r(s t ,a t )]; and a dynamic compensation control sub-unit for performing dynamic compensation control of the system according to the optimal strategy;

[0132] where π(a|s) represents the probability distribution of selecting action a in state s, π * (a|s) represents the optimal strategy, argmax π represents finding the strategy π that maximizes the expected return, and E represents the expectation operator.

[0133] Here, through the autonomous decision-making ability of DRL, precise collaborative control of the multi-energy heating system within a short time domain can be achieved, providing a reliable execution basis for long / medium-term optimization. This DRL framework successfully solves the problem of insufficient adaptability of traditional control methods in fast time-varying and multi-disturbance scenarios, forming a complete closed-loop with long-term time-scale prediction and medium-term time-scale optimization.

[0134] The system constructs a cross-scale information interaction mechanism and a time dimension decoupling and hierarchical optimization mechanism by dividing the time scale.

[0135] The cross-scale information interaction mechanism is as follows:

[0136] Prediction-optimization closed-loop: The long-term prediction results provide constraint boundaries for medium-term optimization, and the medium-term optimization decision serves as a reference trajectory for short-term control.

[0137] Error backpropagation correction: The real-time operation data at the short-term time scale layer is fed back to the medium-term layer through the digital twin mirror model to update the rolling optimization parameters.

[0138] The time dimension decoupling and hierarchical optimization mechanism is as follows:

[0139] For the long-term time scale: An LSTM-Transformer hybrid model is used to predict the output of renewable energy and the heat load demand, and meteorological data is fused through the attention mechanism to establish an energy demand baseline and an equipment start-stop baseline.

[0140] For the medium-term time scale: The NSGA-III multi-objective optimization algorithm is applied to dynamically adjust the energy ratio to balance the objectives of economy (energy consumption cost), reliability (heat network pressure fluctuation), and environmental protection (carbon emissions).

[0141] For the short-term time scale: Minute-level dynamic compensation is achieved based on deep reinforcement learning (DRL). For example, the rotation speed of the heat pump and the opening degree of the heat storage tank valve are adjusted in real time through the Actor-Critic network.

[0142] As a preferred technical solution, the intelligent scheduling system for multi - energy collaborative heating further includes an energy priority matrix construction subunit and a scenario matching subunit.

[0143] The energy priority matrix construction subunit is used to construct an energy priority matrix based on game theory, and the scenario matching subunit is used to match different scenarios through transfer learning according to the constructed energy priority matrix based on game theory.

[0144] Specifically, the energy priority matrix based on game theory obtains the scoring function of energy equipment Adjusts the weight coefficient in the scoring function based on the scoring parameters to match different scenarios.

[0145] Among them, P avail represents the actual available power of the i - th type of energy equipment in the current time period, which can be obtained by collecting the output of renewable energy through a real - time monitoring system. P max represents the upper limit of the installed capacity of the i - th type of energy equipment, which is determined by the rated power of the equipment and the system topology structure, such as the maximum heating capacity of a ground - source heat pump unit. C unit represents the current unit energy cost (yuan / kWh or yuan / GJ) of the i - th type of energy equipment. C base represents the preset benchmark energy cost (yuan / kWh or yuan / GJ), which usually takes the regional average energy cost or the policy - guided price and is normalized for calculation. σ predict represents the standard deviation of the output prediction fluctuation of the i - th type of energy equipment, which can be calculated based on the statistic of historical prediction errors (such as the output prediction standard deviation based on a Bayesian neural network), reflecting the stability of energy supply. For example, the σ of photovoltaic prediction predict can reach 20% - 30% of the installed capacity, and the σ of geothermal energy predict is usually less than 5%.

[0146] Through the said scoring function Quantitatively evaluates the availability, economy and stability of each energy, and can provide a decision - making basis for the scheduling of multi - energy collaborative heating.

[0147] To sum up, the intelligent scheduling system for multi - energy collaborative heating can perform collaborative optimization of multi - energy heating on multiple time scales, improve the overall scheduling effect, and enhance the robustness and generalization ability of the system by constructing a multi - time - scale division mechanism, processing the time scale, dividing it into long - term, medium - term and short - term time scales, and applying different processing strategies for different time scales.

[0148] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0149] The above-described embodiments only express several implementation manners of the present invention, and the description is relatively specific and detailed, but it cannot be understood as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the invention patent should be subject to the appended claims.

Claims

1. An intelligent scheduling method for multi-energy collaborative heating, characterized in that: The intelligent scheduling method for multi-energy coordinated heating comprises the following steps: Construct a multi-time scale division mechanism to process the time scale and divide it into long-term time scale, medium-term time scale and short-term time scale; For the long-term time scale, a hybrid model based on LSTM-Transfermer is constructed to predict the output of renewable energy and heat load demand; For the medium-term time scale, dynamic rolling optimization is adopted to achieve multi-objective optimization through NSGA-Ⅲ algorithm; For the short-term time scale, a deep reinforcement learning (DRL) model is constructed, and dynamic compensation control is performed through the deep reinforcement learning (DRL) model.

2. The intelligent scheduling method for multi-energy coordinated heating according to claim 1, characterized in that: The specific method of building a LSTM-Transformer hybrid model to predict renewable energy output and heat load demand includes the following steps: Obtain the time window feature matrix of the input sequence and extract local time series features through a bidirectional LSTM network; Obtain a meteorological feature matrix, and dynamically fuse meteorological influences through a cross-attention layer based on the obtained meteorological feature matrix; Capturing correlations across time steps through Transformer; Based on the fused meteorological impacts and the correlation of the captured time steps, the forecast values ​​of renewable energy output and heat load demand are obtained.

3. The intelligent scheduling method for multi-energy coordinated heating according to claim 2, characterized in that: The specific method of implementing multi-objective optimization by using dynamic rolling optimization and NSGA-Ⅲ algorithm includes the following steps: Get energy cost C energy 、System loss C loss And user temperature deviation ΔT user , construct the target optimization function min{C energy ,C loss ,ΔT user }; Get the total energy equipment supply ΣE i And the total system heat load demand Q load , according to the total supply of energy equipment and the total heat load demand of the system, the constraint condition s.tΣE is obtained i ≥Q load ; Among them, E i Represents the output energy of the i-th type of energy equipment.

4. The intelligent scheduling method for multi-energy coordinated heating according to claim 3, characterized in that: The specific method of constructing a deep reinforcement learning DRL model and performing dynamic compensation control through the deep reinforcement learning DRL model includes the following steps: Get the state variable s of the heat pump system t and the action variable a t , obtain the reward function based on the state variables and action variables; Get the discount factor, and get the optimal strategy based on the discount factor and reward function; Perform dynamic compensation control of the system according to the optimal strategy.

5. The intelligent scheduling method for multi-energy coordinated heating according to claim 4, characterized in that: The intelligent scheduling method for multi-energy coordinated heating also includes the following steps: Constructing an energy priority matrix based on game theory; According to the constructed game theory-based energy priority matrix, different scenarios are matched through transfer learning.

6. An intelligent scheduling system for multi-energy collaborative heating, used to implement the intelligent scheduling method for multi-energy collaborative heating as described in any one of claims 1 to 5, characterized in that: The intelligent dispatching system for multi-energy coordinated heating includes a time scale division module, which is used to construct a multi-time scale division mechanism, process the time scale, and divide it into a long-term time scale, a medium-term time scale, and a short-term time scale. The time scale division module includes: The long-term time scale unit is used to build a LSTM-Transfermer hybrid model to predict the output of renewable energy and heat load demand for the long-term time scale; The medium-term time scale unit is used to achieve multi-objective optimization by using dynamic rolling optimization and NSGA-Ⅲ algorithm for medium-term time scale; The short-term time scale unit is used to construct a deep reinforcement learning (DRL) model for the short-term time scale and perform dynamic compensation control through the deep reinforcement learning (DRL) model.

7. The intelligent scheduling system for multi-energy coordinated heating according to claim 6, characterized in that: Long-term time scale units include: The local time series feature acquisition subunit is used to obtain the time window feature matrix of the input sequence and extract the local time series features through the bidirectional LSTM network; The meteorological impact acquisition subunit is used to acquire the meteorological feature matrix and dynamically fuse the meteorological impact through the cross attention layer according to the acquired meteorological feature matrix; The correlation acquisition subunit is used to capture the correlation across time steps through the Transformer; The prediction subunit is used to obtain the predicted values ​​of renewable energy output and heat load demand based on the fused meteorological impact and the correlation of the captured time steps.

8. The intelligent scheduling system for multi-energy coordinated heating according to claim 7, characterized in that: Medium-term time scale units include: The target optimization function obtains the subunit, which is used to obtain the energy cost C energy 、System loss C loss And user temperature deviation ΔT user , construct the target optimization function min{C energy ,C loss ,ΔT user }; Constraint acquisition subunit, used to obtain the total supply of energy equipment ΣE i And the total system heat load demand Q load , according to the total supply of energy equipment and the total heat load demand of the system, obtain the constraint condition st∑E i ≥Q load ; Among them, E i Represents the output energy of the i-th type of energy equipment.

9. The intelligent scheduling system for multi-energy coordinated heating according to claim 8, characterized in that: Short-term time scale units include: A reward function acquisition subunit is used to acquire state variables and action variables of the heat pump system, and acquire a reward function according to the state variables and action variables; The optimal strategy acquisition subunit is used to obtain the discount factor and obtain the optimal strategy based on the discount factor and the reward function; The dynamic compensation control subunit is used to perform dynamic compensation control of the system according to the optimal strategy.

10. The intelligent scheduling system for multi-energy coordinated heating according to claim 9, characterized in that: The intelligent dispatching system for multi-energy coordinated heating also includes: Energy priority matrix construction subunit, used to construct the energy priority matrix based on game theory; The scenario matching subunit is used to match different scenarios through transfer learning according to the constructed game theory-based energy priority matrix.

Images