Load prediction and optimal scheduling method and system for multi-energy-storage thermal power generating unit

Abstract

The invention discloses a multi-energy-storage thermal power generating unit load prediction and optimal scheduling method and system, and relates to the technical field of multi-energy-storage thermal power generating units, and the method comprises the steps: collecting the operation parameters and external environment parameters of a thermal power generating unit in real time, and constructing a real-time state parameter matrix; constructing a load prediction model based on the historical state parameter matrix, importing the real-time state parameter matrix into the load prediction model, outputting a load trend prediction curve, and triggering an early warning signal through a secondary discrimination mechanism; identifying a load disturbance value based on the load trend prediction curve, obtaining a load disturbance sequence, and decoupling the load disturbance sequence into a plurality of components; inputting the plurality of vectors into a preset decision network, dynamically correcting a constraint condition built in the decision network in combination with the early warning signal, introducing an improved dragonfly algorithm for optimization iteration, and generating an optimization scheduling instruction; according to the method, the adaptability of optimal scheduling and high-precision prediction of the load trend are improved.

Description

Technical Field

[0001] This invention relates to the field of multi-energy storage thermal power unit technology, specifically to a method and system for load prediction and optimized scheduling of multi-energy storage thermal power units. Background Technology

[0002] Thermal power units, as the core foundation of the power system, play a crucial role in ensuring the safe and stable operation of the power grid. Especially with the high proportion of renewable energy integration, their primary function has gradually shifted from bearing base load to providing deep peak shaving and rapid frequency regulation to mitigate the random fluctuations of wind and solar power generation. Although existing technical solutions have begun to widely incorporate artificial intelligence and big data technologies, such as using long short-term memory networks for load forecasting and particle swarm optimization for scheduling optimization, these existing applications still have significant limitations. On the one hand, existing technologies typically input historical data directly into neural networks to output load forecasts. Based on these forecasts, load scheduling and control of thermal power units and their associated energy storage systems are performed. For example, frequency deviation and SOC changes are often used as adjustment criteria to regulate energy storage output, without considering load disturbances. This makes it difficult to meet the refined control requirements of multi-thermal-storage combined systems under complex operating conditions. Furthermore, most scheduling technologies adopt conventional AGC command response modes, passively adjusting boiler fuel quantity, turbine control valves, or battery charging and discharging power based solely on the difference between predicted and actual values ​​using PID logic or simple threshold rules. On the other hand, existing technologies lack high-precision prediction of ultra-short-term load change trends, making it impossible to pre-set energy levels before disturbances occur. Additionally, internal constraints in scheduling control, such as ramp rate limits, are typically static and cannot be dynamically corrected based on real-time energy storage charge status. Moreover, the optimization algorithms used often get trapped in local optima, resulting in poor control adaptability under complex and variable operating conditions, making it difficult to achieve optimal global energy efficiency for the entire system. Summary of the Invention

[0003] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method and system for load forecasting and optimized scheduling of multi-energy storage thermal power units. By collecting operating parameters of thermal power units and external environmental parameters, a real-time state parameter matrix is ​​constructed. Based on the historical state parameter matrix, a load forecasting model is built, and a two-level discrimination mechanism based on amplitude and duration is established. This enables high-precision prediction of future load trend changes and timely early warning of gap sequences. Through an improved dragonfly algorithm, combined with the early warning signal, the built-in constraints of the decision network are dynamically corrected, thereby improving the algorithm performance of the decision network and enabling dynamic adjustment of scheduling instructions. This solves the problems mentioned in the background technology.

[0004] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, this application provides a method for load prediction and optimized scheduling of multi-energy storage thermal power units, the method comprising: real-time acquisition of thermal power unit operating parameters and external environmental parameters, and construction of a real-time state parameter matrix; A load forecasting model is constructed based on the historical state parameter matrix. The real-time state parameter matrix is ​​imported into the load forecasting model, the load trend forecast curve is output, and an early warning signal is triggered through a two-level discrimination mechanism. Based on the load trend prediction curve, load disturbance values ​​are identified, load disturbance sequences are obtained, and decoupled into several components. Several vectors are input into a preset decision network, and the constraints built into the decision network are dynamically corrected in combination with the early warning signal. An improved dragonfly algorithm is introduced to iterate and generate optimized scheduling instructions.

[0005] Furthermore, the operating parameters include the real-time output of the thermal power unit, the grid frequency deviation, the ACG load command, and the SOC state of charge, while the external environmental parameters include wind speed, temperature, and irradiance.

[0006] Furthermore, a load forecasting model is constructed based on the historical state parameter matrix, including: The load forecasting model is built on a graph convolutional network architecture, including a working condition feature recognition layer, a load trend evolution layer, and a power gap triggering layer. Operating condition feature identification layer: Establish a feature matrix containing typical load conditions, and determine the current load condition type by matching the real-time state parameter matrix with the feature matrix; where the matching degree is calculated using cosine similarity. Load trend evolution layer: Based on the current load condition type, a multi-level prediction algorithm is adopted to adaptively match the preset combined model hyperparameters and establish the first, second and third prediction sub-models respectively. The weighted outputs of each sub-model are spliced ​​together to generate a load trend prediction curve covering future time. Power Shortage Triggering Layer: The baseline output of thermal power units is used to construct the basic output trajectory, calculate the time-domain difference between this trajectory and the load trend prediction curve, and aggregate them into a shortage sequence; the amplitude and duration of the power shortage sequence are extracted, and a secondary early warning system is set. The threshold is determined through a two-level discrimination mechanism. Cases with amplitude values ​​greater than the first warning threshold are selected to trigger the first warning signal. At the same time, the product of amplitude and duration is calculated, and cases with product results greater than the second warning threshold are selected to trigger the second warning signal.

[0007] Furthermore, a feature matrix containing typical load conditions is established, including: Historical state parameter matrices covering periods of stable load, periods of sudden load change, and periods of sudden load drop are collected. The features of the historical state parameter matrix are denoised to separate low-frequency trend terms and high-frequency random terms, and the denoised feature matrix is ​​reconstructed. The correlation between the denoised wind speed, irradiance, temperature and ACG load command fluctuations is analyzed using Pearson correlation coefficient. Key parameter features are screened by the maximum mutual information coefficient.

[0008] Furthermore, the decoupling is divided into several components, including: The identified load disturbance sequence is iteratively decomposed using variational mode decomposition to obtain several intrinsic mode functions; Based on the intrinsic mode function, instantaneous amplitude and instantaneous frequency are extracted to construct a two-dimensional feature space, and a feature trajectory curve is formed by fitting. Based on the feature trajectory curve, the zero-crossing point and its direction of change are identified, and several monotonic fluctuation intervals are divided. Cluster analysis of monotonic fluctuation intervals identifies several components, including low-frequency, mid-frequency, and high-frequency components.

[0009] Furthermore, the preset decision network includes a low-frequency reference branch, a mid-frequency buffer branch, and a high-frequency response branch. The low-frequency reference branch has a built-in first constraint, the mid-frequency buffer branch has a built-in second constraint, and the high-frequency response branch has a built-in third constraint.

[0010] Furthermore, the constraints built into the decision network are dynamically adjusted based on the early warning signals, including: A multidimensional constraint parameter space is constructed based on the low-frequency reference branch, the mid-frequency buffer branch, and the high-frequency response branch. The multidimensional constraint parameter space is orthogonally constructed by the first constraint, the second constraint, and the third constraint. At the same time, the boundary parameters of each constraint condition are extracted, including the boundary interval, the upper boundary limit, and the lower boundary limit. In conjunction with the early warning signal, the multidimensional constraint parameter space is dynamically reconstructed. Under the condition of responding to the first early warning signal, a first switching strategy is executed for the boundary parameters of the first constraint, and a second switching strategy is executed for the boundary parameters of the third constraint; under the condition of responding to the second early warning signal, a third switching strategy is executed for the boundary parameters of the second constraint, and a fourth switching strategy is executed for the boundary parameters of the third constraint. Latin hypercube sampling is used to perform hierarchical sampling in the multidimensional constraint parameter space to generate several uniformly distributed sample points; the sample points are combined to generate a correction vector; the correction vector contains the combination of boundary parameters corresponding to different constraint conditions.

[0011] Furthermore, the first constraint includes fuel ramp rate and unit output amplitude, the second constraint includes temperature regulation dead zone and valve operating frequency, and the third constraint includes charging and discharging power and charging and discharging trigger boundary.

[0012] Furthermore, an improved dragonfly algorithm is introduced for optimization iteration, including: The correction vector is used as the initial position vector of the dragonfly population in the Dragonfly Algorithm. Before each iteration, a state deduction operation is performed, including: updating the constraints in the decision network based on the correction vector; simultaneously, retrieving several currently obtained components and establishing the correlation between the updated constraints and the components; and extracting the first evaluation function and the second evaluation function based on the correlation. The first evaluation function is as follows: extract the amplitude of the low-frequency component and calculate the ratio with the corresponding first constraint, and mark it as the first index; extract the amplitude of the mid-frequency component and calculate the difference with the corresponding second constraint, and mark it as the second index; construct the first evaluation function based on the first index and the second index. The second evaluation function is as follows: extract the amplitude of the high-frequency component and calculate the difference with the corresponding third constraint, and mark it as the third index; extract the frequency of change of the mid-frequency component and calculate the deviation from the second constraint, and mark it as the fourth index; construct the second evaluation function based on the third index and the fourth index. In each iteration, the position vector is updated by maximizing the first evaluation function and minimizing the second evaluation function until the maximum number of iterations is satisfied.

[0013] Secondly, this application provides a load prediction and optimization scheduling system for multi-energy storage thermal power units. The system includes: a data acquisition module, which collects the operating parameters of the thermal power units and external environmental parameters in real time and constructs a real-time status parameter matrix; The load forecasting module constructs a load forecasting model based on the historical state parameter matrix, imports the real-time state parameter matrix into the load forecasting model, outputs a load trend forecasting curve, and triggers an early warning signal through a two-level discrimination mechanism. The optimized scheduling module identifies load disturbance values ​​based on the load trend prediction curve, obtains the load disturbance sequence, and decouples it into several components. Several vectors are input into a preset decision network, and the constraints built into the decision network are dynamically corrected in combination with the early warning signal. An improved dragonfly algorithm is introduced to iterate and generate optimized scheduling instructions.

[0014] (III) Beneficial Effects This invention provides a method and system for load prediction and optimized scheduling of multi-energy storage thermal power units, which has the following beneficial effects: 1. This invention constructs a high-dimensional real-time state parameter matrix by collecting operating parameters of thermal power units and external environmental parameters; the synchronous sampling mechanism triggered by a unified clock ensures the alignment of data in the time domain, eliminates the analysis error caused by data asynchrony, and provides a high-fidelity data foundation for the input of subsequent load forecasting models; 2. This invention constructs a load prediction model based on a historical state parameter matrix. Through a working condition feature identification layer, a load trend evolution layer, and a power gap triggering layer, it achieves high-precision prediction of load trend changes in the future, especially the prediction of ultra-short-term load trends. Combined with a real-time state parameter matrix, a two-level discrimination mechanism based on amplitude and duration is established to achieve timely early warning of gap sequences. 3. This invention improves the performance of the decision network through the improved Dragonfly algorithm, enabling dynamic adjustment of scheduling instructions. Based on the correlation between low-frequency, mid-frequency, and high-frequency components and their corresponding constraints, a first evaluation function and a second evaluation function are established. By maximizing the first evaluation function and minimizing the second evaluation function, the basic scheduling instructions are dynamically adjusted to generate more accurate optimized scheduling instructions, thereby improving energy utilization. 4. This invention uses a pre-defined decision network and combines it with early warning signals to dynamically correct the constraints built into the decision network, making the scheduling process more centralized and easier to manage, enhancing the stability and generalization ability of the model, improving computational efficiency, and making the system easier to adjust and optimize. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the steps of a load forecasting and optimized scheduling method according to an exemplary embodiment; Figure 2 This is a schematic diagram of the framework of a load forecasting model according to an exemplary embodiment; Figure 3 This is a schematic diagram of a load forecasting and optimization scheduling system according to an exemplary embodiment. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1: This invention provides a method for load prediction and optimized scheduling of multi-energy storage thermal power units; Figure 1 This is a schematic diagram illustrating the steps of a load forecasting and optimized scheduling method according to an exemplary embodiment; Figure 2 This is a schematic diagram illustrating the framework of a load forecasting model according to an exemplary embodiment; please refer to... Figures 1 to 2 The method includes the following steps: S1: Real-time acquisition of thermal power unit operating parameters and external environmental parameters to construct a real-time status parameter matrix; among which, operating parameters include real-time output of thermal power unit, grid frequency, ACG load command and SOC state of charge, and external environmental parameters include wind speed, temperature and irradiance. Power transmitter arrays are deployed at the generator outlet bus and grid connection point of thermal power units to collect operating parameters. Specifically, this includes: installing Hall effect power transmitters at the three-phase stator outlet of the generator, for example, with a sampling rate of 1kHz, to sense voltage and current signals in real time and calculate them into the real-time output of the thermal power unit; installing high-precision frequency transmitters on the high-voltage side bus at the grid connection point to capture the zero-crossing point of the grid voltage and measure the grid frequency in real time; and simultaneously, capturing digital pulse signals issued by the grid dispatch center in real time through the hard-wired signal port of the unit control cabinet and parsing them into AGC load commands. On the DC bus and individual battery terminals of the hybrid energy storage device, a battery state sensor array is deployed to collect the SOC state of charge. Specifically, this includes: using a shunt connected in series in the main circuit to measure the charging and discharging current in real time, using a voltage acquisition probe to measure the terminal voltage, transmitting the collected voltage, current, and temperature raw data to the local computing unit, and calculating the SOC state of charge in real time by combining the open-circuit voltage method and the ampere-hour integration method. In the open area of ​​the thermal power plant, next to the wind measurement tower and photovoltaic array, a physical environment monitoring sensor group is deployed to synchronously collect external environmental parameters. Specifically, this includes: deploying ultrasonic wind speed sensors to measure air flow speed by utilizing the time difference of ultrasonic waves propagating in the air and outputting real-time wind speed; deploying industrial-grade platinum resistance thermometers to measure air thermodynamic temperature by utilizing the characteristic of conductor resistance changing with temperature and outputting ambient temperature; and deploying photoelectric total radiation sensors to convert solar radiation energy into electrical signals using photosensitive elements and outputting irradiance. The analog and digital signals collected by the above-mentioned sensing devices are converted from analog to digital and synchronized in time to construct a real-time state parameter matrix; specifically, a unified clock-triggered sampling method is adopted to synchronously latch the data of all sensors at a fixed sampling frequency; The significance of the above analysis lies in the following: by rationally deploying underlying physical sensors in the electrical circuits of thermal power units, energy storage devices, and plant environmental monitoring points, it is possible to accurately collect multi-source heterogeneous data such as unit output, grid frequency, SOC status, and wind and solar environment, and construct a high-dimensional real-time status parameter matrix; the unified clock-triggered synchronous sampling mechanism ensures data alignment in the time domain, eliminates analysis errors caused by data asynchrony, and provides a high-fidelity data foundation for the input of subsequent load forecasting models.

[0018] S2: Construct a load forecasting model based on the historical state parameter matrix, import the real-time state parameter matrix into the load forecasting model, and output the load trend forecasting curve; A load forecasting model is constructed based on the historical state parameter matrix, including: The full-cycle state parameter matrix covering the load stability period, load change period, and load drop period over the past three years was collected. Mean filtering was applied to the collected state parameter matrix to remove outliers caused by sensor malfunctions, and the maximum-minimum normalization method was used to map all parameters to the [0, 1] interval to eliminate the influence of different dimensions. Load stability period: refers to a condition where the external power grid load demand is relatively stable, the AGC load command received by the thermal power unit shows a smooth trend, and the power grid frequency deviation remains within the regulation dead zone. The unit can meet its output requirements without frequent adjustments to the turbine valve opening. Load change period: refers to a condition where significant changes in external environmental parameters such as wind speed and irradiance cause drastic fluctuations in the output of new energy units in the region, thus forcing the AGC load command received by the thermal power unit to exhibit large-amplitude, high-frequency oscillations. Load drop period: refers to a condition where the thermal power unit operates in a low-load range far from its rated operating condition to cooperate with the power grid's deep peak shaving and absorption of new energy. Under this condition, the boiler combustion stability of the thermal power unit decreases, and due to the inertia of the thermal system, its actual physical ramp-up rate is significantly lower than the response rate under rated operating conditions. The normalized historical state parameter matrix is ​​subjected to discrete wavelet transform algorithm to perform column vector decomposition, that is, the time series of each dimension parameter is decomposed into multiple scales to separate the low-frequency trend term representing long-term regularity and the high-frequency random term representing noise. Discrete wavelet transform is then performed on the separated low-frequency trend term representing long-term regularity and the high-frequency random term representing noise, and the denoised feature matrix is ​​reconstructed. The correlation between the denoised wind speed, irradiance, temperature and ACG load command fluctuation is analyzed using Pearson correlation coefficient, and key parameter features are screened by maximum mutual information coefficient. At the same time, the maximum mutual information coefficient is calculated to screen key parameter features, which are used as input variables for matching degree judgment in the subsequent working condition feature recognition layer. The load forecasting model is built on a graph convolutional network architecture, including a working condition feature recognition layer, a load trend evolution layer, and a power gap triggering layer. Operating condition feature identification layer: Establish a feature matrix containing the above three typical load conditions: Determine the current load condition type by matching the real-time state parameter matrix with the feature matrix; where the matching degree is calculated using cosine similarity. Load Trend Evolution Layer: Based on the current load condition type, a multi-level prediction algorithm is adopted, adaptively matching preset combined model hyperparameters to establish the first, second, and third prediction sub-models respectively. It should be noted that the first prediction sub-model represents a super-short-term prediction model at the second level, e.g., the next 30 seconds; the second prediction sub-model represents a short-term prediction model at the minute level, e.g., the next 15 minutes; and the third prediction sub-model represents a day-ahead prediction model at the hour level, e.g., 12 hours. The first prediction sub-model is configured to process the real-time state parameter matrix of high-frequency sampling. It adopts a temporal convolutional network containing two one-dimensional convolutional layers and a max-pooling layer. It utilizes its dilated causal convolutional structure to expand the receptive field, capture the instantaneous mutation features and random noise components of the load sequence, and outputs the load disturbance sequence with a prediction time window of the next 30 seconds. The second prediction sub-model is configured to process the state parameter matrix of minute-level resampling. It adopts a bidirectional gated recurrent architecture and uses the forward and reverse gating mechanism of BiGRU to capture the inertial trend and ramp-up features of the load sequence on the intraday time scale. It outputs the load disturbance sequence with a prediction time window of the next 15 minutes. The third prediction sub-model is configured to process the real-time state parameter matrix of hour-level resampling. It adopts a Transformer encoder based on a self-attention mechanism to process the state parameter matrix of daily resampling, extract the weekly periodicity and long-term trend of the load, and output the load disturbance sequence for the next 12 hours. The specific parameters mentioned above are all reference values, and the actual values ​​are determined according to the actual situation. The weighted outputs of each sub-model are concatenated and processed. The lower and upper bounds of the prediction confidence are calculated through the residual distribution to form an envelope curve and generate a load trend prediction curve covering future time. In addition, during the prediction phase, the probability prediction results with confidence intervals are output. When the actual output trend prediction value exceeds the confidence limit, online incremental learning is triggered to update the model parameters using the latest data. Power Shortage Triggering Layer: The baseline output of the thermal power unit is used to construct the basic output trajectory, calculate the time-domain difference between the baseline output and the load trend prediction curve, and aggregate the time-domain difference into a shortage sequence; the amplitude and duration of the power shortage sequence are extracted, and a two-level early warning threshold is set, including a first early warning threshold and a second early warning threshold; it should be noted that the baseline output of the thermal power unit is a value built into the system, and a parallel line is drawn on the horizontal axis of this value, which is the basic output trajectory; in this embodiment, the horizontal axis of the load trend prediction curve is set as a time series; Through a two-level discrimination mechanism: Level 1 discrimination: The amplitude is compared with the first warning threshold. If the amplitude is less than or equal to the first warning threshold, it is determined to be a normal control state and no warning signal is triggered. If the amplitude is greater than the first warning threshold, it is determined to be a power over-limit, triggering the first warning signal and performing level 2 discrimination. Secondary discrimination: Under the condition of triggering the first warning signal, calculate the product of amplitude and duration, and compare the product result with the second warning threshold: if the product result is less than or equal to the second warning threshold, it is determined to be a short-term impact state, and the first warning signal is maintained; if the product result is greater than the second warning threshold, it is determined to be a continuous overload state, and the second warning signal is triggered. By comparing the accuracy and false alarm rate under different thresholds through cross-validation, the threshold with the best accuracy is selected as the first warning threshold and the second warning threshold. The significance of the above analysis lies in the following: by constructing a load prediction model based on the historical state parameter matrix, and through the operating condition feature identification layer, load trend evolution layer, and power gap triggering layer, the nonlinear characteristics between thermal power unit output, SOC (State of Charge), and external environmental parameters can be deeply explored. By combining the real-time state parameter matrix, a two-level discrimination mechanism based on amplitude and duration is established, which realizes timely early warning of gap sequences and high-precision prediction of load trend changes in the future.

[0019] S3: Identify load disturbance values ​​based on load trend prediction curves, obtain load disturbance sequences, and decouple them into several components; input several vectors into a preset decision network, dynamically correct the constraints built into the decision network in conjunction with early warning signals, introduce an improved dragonfly algorithm for optimization iteration, and generate optimized scheduling instructions. The process of identifying load disturbance values ​​based on load trend prediction curves and obtaining load disturbance sequences includes: extracting the value corresponding to the prediction point at the current moment as the trend prediction value based on the load trend prediction curve; extracting the current AGC load command and grid frequency deviation from the real-time state parameter matrix, and parsing the current real-time load value based on the AGC load command; calculating the difference between the trend prediction value and the real-time load value and defining it as the deviation component; calculating the power component representing the grid frequency regulation demand based on the grid frequency deviation and a preset frequency regulation droop coefficient product operation; wherein the preset frequency regulation droop coefficient is greater than 0; performing algebraic superposition of the deviation component and the power component to generate the instantaneous load disturbance value at the current moment; establishing a sliding time window of preset length; storing the instantaneous load disturbance values ​​at the current moment and several consecutive historical sampling moments in chronological order into the sliding time window to construct a load disturbance sequence, which serves as the input data for subsequent decoupling. Decoupling into several components, including: Variational mode decomposition is used to iteratively decompose the identified load disturbance sequence to obtain several eigenmode functions with different center frequencies; Hilbert transform is performed on each intrinsic mode function to construct an analytic signal in the complex domain, and the instantaneous amplitude and instantaneous frequency at each moment are calculated based on the analytic signal. A two-dimensional characteristic phase space is constructed with the instantaneous frequency as the abscissa and the instantaneous amplitude as the ordinate. The load disturbance values ​​at each moment are mapped to characteristic points in the characteristic space, and the characteristic points at consecutive moments are connected to form a characteristic evolution trajectory. The first derivative of the feature evolution trajectory in the amplitude dimension is calculated based on the feature evolution trajectory. The zero-crossing point where the first derivative changes from positive to negative or from negative to positive is identified. The zero-crossing point is used as the dividing node to truncate the sequence of intrinsic mode functions into several monotonic fluctuation intervals. Cluster analysis was performed on all monotonic fluctuation intervals to identify several components, including low-frequency, mid-frequency, and high-frequency components. Specifically, the following steps were taken: K-Means clustering was applied to the feature points of all monotonic fluctuation intervals in the two-dimensional feature model. The preset number of clusters was 3. Based on the principle of minimizing the intra-cluster variance, the centroid positions of each cluster were iteratively updated until the algorithm converged, resulting in three independent feature point clusters. The Euclidean distance from the centroid of each of the three feature point clusters to the origin of the phase space coordinate system was calculated and sorted in ascending order of distance. The feature point cluster with the smallest centroid distance was labeled as the steady-state cluster, and its monotonic fluctuation intervals were grouped into low-frequency components. The feature point cluster with the middle centroid distance was labeled as the transition cluster, and its monotonic fluctuation intervals were grouped into mid-frequency components. The feature point cluster with the largest centroid distance was labeled as the transient cluster, and its monotonic fluctuation intervals were grouped into high-frequency components. The pre-defined decision network includes a low-frequency reference branch, a mid-frequency buffer branch, and a high-frequency response branch. Low-frequency reference branch: Built-in first constraint, which includes fuel ramp rate and unit output amplitude; Intermediate frequency buffer branch: Built-in second constraint, which includes temperature regulation dead zone and valve operating frequency; High-frequency response branch: Built-in third constraint, which includes charge / discharge power and charge / discharge trigger boundary; By inputting several components into a preset decision network, basic scheduling instructions are generated, including boiler fuel quantity instructions, heating regulating valve opening instructions, and battery charging and discharging instructions: low-frequency components are converted into boiler fuel quantity instructions; medium-frequency components are converted into opening instructions for the turbine connecting pipe butterfly valve and the heating extraction steam butterfly valve, and the virtual energy storage of the heating network is used to smooth out fluctuations; high-frequency components are converted into electrochemical energy storage charging and discharging instructions. The significance of the above analysis is that by identifying load disturbance values ​​based on load trend prediction curves and decoupling the load disturbance sequence into low-frequency, medium-frequency, and high-frequency components, the coupling complexity of multi-source collaborative control is reduced to a certain extent. By accurately matching components of different frequencies to the boundary parameter ranges of thermal power, heating networks, and energy storage, the system's tracking accuracy for variable loads and the safety of equipment operation are improved. The constraints built into the decision network are dynamically adjusted based on the early warning signals, including: A multi-dimensional constraint parameter space is constructed based on the low-frequency reference branch, the mid-frequency buffer branch, and the high-frequency response branch; wherein, the multi-dimensional constraint parameter space is orthogonally constructed by the first constraint, the second constraint, and the third constraint; At the same time, the boundary parameters of each constraint are extracted, including the boundary interval, the upper boundary limit, and the lower boundary limit; Based on the early warning signal, the dynamic reconstruction of the multi-dimensional constraint parameter space, in response to the first early warning signal, implements a first switching strategy for the boundary parameters of the first constraint, including: increasing the upper limit of the fuel ramp rate boundary and relaxing the unit output amplitude boundary range. For example, replacing the upper limit of the fuel ramp rate with the physical limit ramp rate, and locking the feasible domain of the unit output amplitude within the maximum rated output range, forcing the thermal power unit to operate in the rated high load range, i.e., resetting the lower limit of the output amplitude range boundary. Simultaneously, a second switching strategy is implemented for the boundary parameters of the third constraint, including: increasing the upper limit of the charging and discharging power boundary and relaxing the charging and discharging trigger boundary. For example, increasing the upper limit of the charging and discharging power boundary from the rated operating power threshold to the short-term overload power threshold, and relaxing the charging and discharging trigger boundary. The discharge dead zone is limited, and the trigger probability range of the discharge logic is expanded. Under the condition of responding to the second warning signal: the third switching strategy is implemented on the boundary range of the second constraint, including: narrowing the boundary range of the temperature regulation dead zone and the boundary range of the valve action frequency, for example: reducing the range of the temperature regulation dead zone value and the range of the valve action frequency, and performing a compression process on the minimum time interval of valve action; at the same time, the fourth switching strategy is implemented on the lower limit of the boundary of the third constraint, including: raising the lower limit of the charge and discharge trigger boundary, for example: raising the discharge cutoff threshold in the charge and discharge trigger boundary from the deep discharge threshold to the shallow discharge threshold, while raising the lower limit of the boundary, keeping the upper limit of the charging direction boundary unchanged, thereby limiting the continuous discharge behavior of the hybrid energy storage system and encouraging it to perform charging behavior; Latin hypercube sampling is employed, with stratified sampling within the multidimensional constraint parameter space to generate several uniformly distributed sample points, specifically including: Based on the dynamically reconstructed multidimensional constraint parameter space, the range of boundary parameter values ​​for each constraint is divided into M non-overlapping, equally probable intervals, where M is greater than or equal to 1. For example, assuming M is 5, the fuel ramp rate in the first constraint is uniformly divided into 5 continuous sub-intervals; the valve actuation frequency in the second constraint is divided into 5 corresponding integer layers. A sample point is randomly selected from each divided interval to ensure that there is only one sample in each row and column, and numerical mapping is performed. For continuous parameters, including fuel ramp rate, unit output amplitude, temperature regulation dead zone, charging and discharging power, and trigger boundary, floating-point precision is retained for the sampled values, for example, retaining two decimal places. For discrete parameters, including valve actuation frequency limit, the nearest integer rounding operation is performed on the sampled normalized values. The above sample points are combined to generate a correction vector, where the correction vector contains combinations of boundary parameters corresponding to different constraint conditions. An improved Dragonfly algorithm is introduced for optimization iteration to generate optimized scheduling instructions, including: The Dragonfly Algorithm is an optimization algorithm developed by mimicking the static predation and dynamic migration behaviors of dragonflies. It has good global search capabilities and convergence speed, and performs well in parameter optimization. In this embodiment, the initial number of dragonflies can be set to 200, the upper and lower bounds of the search range can be defined as 10 and -10 respectively, the inertia factor can be set to 0.9, and the maximum number of iterations can be 300. The initial position and initial velocity of each dragonfly can be set, and the fitness of each dragonfly can be calculated. The global optimal position can be found and recorded. The position and velocity of each dragonfly can be updated according to the mutual attraction and repulsion between individual dragonflies. The fitness of each dragonfly can be recalculated, the global optimal position can be recorded, and the iteration can be started. When the number of iterations reaches 300, the optimal solution can be output. It should be noted that, to enhance the algorithm's local exploitation capabilities and prevent isolated dragonflies from affecting the overall convergence speed, the currently generated correction vector is used as the initial position vector of the dragonfly population in the dragonfly algorithm. Before each iteration, a state deduction operation is performed, including: Based on the correction vector, the boundary parameters of the first, second, and third constraints in the decision network are updated online to complete the parameter reconstruction of the decision network; at the same time, several currently obtained components are retrieved to establish the correlation between the updated constraints and several components. Based on the correlation, the first evaluation function and the second evaluation function are extracted; The first evaluation function is as follows: extract the amplitude of the low-frequency component and calculate the ratio with the corresponding first constraint, that is, calculate the ratio of the amplitude of the low-frequency component to the unit output amplitude boundary, and mark it as the first index; extract the amplitude of the mid-frequency component and calculate the difference with the corresponding second constraint, that is, calculate the difference between the amplitude of the mid-frequency component and the temperature regulation dead zone width, and mark it as the second index; construct the first evaluation function based on the first index and the second index; The second evaluation function is as follows: extract the amplitude of the high-frequency component and calculate the difference with the corresponding third constraint, that is, calculate the difference between the amplitude of the high-frequency component and the charge / discharge power boundary, and mark it as the third index; extract the changing frequency of the mid-frequency component and calculate the deviation from the second constraint, that is, calculate the deviation between the changing frequency of the mid-frequency component and the valve action frequency, and mark it as the fourth index; construct the second evaluation function based on the third index and the fourth index. In each iteration, the position vector is updated by maximizing the first evaluation function and minimizing the second evaluation function until the maximum number of iterations is met, thus obtaining the optimal position vector. The globally optimal position vector is then physically decoded to extract the optimal fuel ramp rate, optimal unit output amplitude, optimal temperature regulation dead zone, optimal charge / discharge power boundary, and optimal charge / discharge trigger boundary. The optimal parameters and several components obtained from the above decoding are then re-introduced into a pre-defined decision network. The currently obtained components are then subjected to amplitude limiting and shaping processing to generate the final optimized scheduling instructions, including: generating optimized boiler fuel quantity instructions, heating regulating valve opening instructions, and battery charge / discharge instructions. For example, the optimal fuel ramp rate is used as a slope constraint, and the optimal unit output amplitude is used as an amplitude constraint. The process involves: constraining the low-frequency components; calculating the algebraic sum of the constrained low-frequency components and the current unit base load, converting it into a boiler fuel quantity control command; using the optimal temperature regulation dead zone as the threshold parameter for the nonlinear dead zone element, and inputting the intermediate-frequency components into this nonlinear dead zone element; if the amplitude of the intermediate-frequency components falls within the dead zone, outputting a hold command; if it exceeds the dead zone range, calculating the excess regulation amount and converting it into a heating regulating valve opening control command; using the optimal charge / discharge power boundary as the output limiting condition, and the optimal charge / discharge trigger boundary as the logic start condition; determining whether the high-frequency components meet the start condition: if they do, performing amplitude limiting on the high-frequency components and converting the limited value into a hybrid energy storage system charge / discharge power command; if they do not meet the condition, generating a standby or forced charging command. The significance of the above analysis lies in the following: By using a pre-defined decision network and dynamically correcting the constraints built into the decision network in conjunction with early warning signals, the scheduling process becomes more centralized and easier to manage, enhancing the stability and generalization ability of the model, improving computational efficiency, and making the system easier to adjust and optimize. Furthermore, the introduction of an improved dragonfly algorithm enhances the algorithmic performance of the decision network, enabling dynamic adjustment of scheduling instructions. Based on the correlation between low-frequency, mid-frequency, and high-frequency components and their corresponding constraints, a first evaluation function and a second evaluation function are established. By maximizing the first evaluation function and minimizing the second evaluation function, the basic scheduling instructions are dynamically adjusted, generating more accurate optimized scheduling instructions and improving energy utilization.

[0020] Example 2: This invention provides a load prediction and optimized scheduling system for multi-energy storage thermal power units; Figure 3 This is a schematic diagram of a load forecasting and optimization scheduling system according to an exemplary embodiment; please refer to... Figure 3 The system includes: a data acquisition module, a load forecasting module, and an optimization scheduling module, and the data acquisition module, load forecasting module, and optimization scheduling module are interconnected. Data acquisition module: Real-time acquisition of thermal power unit operating parameters and external environmental parameters to construct a real-time status parameter matrix; Load forecasting module: Constructs a load forecasting model based on historical state parameter matrix, imports real-time state parameter matrix into the load forecasting model, outputs load trend forecast curve, and triggers early warning signal through a two-level discrimination mechanism; The optimized scheduling module identifies load disturbance values ​​based on the load trend prediction curve, obtains the load disturbance sequence, and decouples it into several components. Several vectors are input into a preset decision network, and the constraints built into the decision network are dynamically corrected in combination with the early warning signal. An improved dragonfly algorithm is introduced to iterate and generate optimized scheduling instructions.

[0021] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0022] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0023] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for load prediction and optimal scheduling of multi-energy storage thermal power units, characterized in that, The method includes: Real-time acquisition of thermal power unit operating parameters and external environmental parameters to construct a real-time status parameter matrix; A load forecasting model is constructed based on the historical state parameter matrix. The real-time state parameter matrix is ​​imported into the load forecasting model, the load trend forecast curve is output, and an early warning signal is triggered through a two-level discrimination mechanism. Based on the load trend prediction curve, load disturbance values ​​are identified, load disturbance sequences are obtained, and decoupled into several components. Several vectors are input into a preset decision network, and the constraints built into the decision network are dynamically corrected in combination with the early warning signal. An improved dragonfly algorithm is introduced to iterate and generate optimized scheduling instructions.

2. The method for load prediction and optimized scheduling of multi-energy storage thermal power units according to claim 1, characterized in that, The operating parameters include real-time output of thermal power units, grid frequency deviation, ACG load command and SOC state of charge, and external environmental parameters include wind speed, temperature and irradiance.

3. The method for load prediction and optimized scheduling of multi-energy storage thermal power units according to claim 2, characterized in that, The load forecasting model constructed based on the historical state parameter matrix includes: The load forecasting model is built on a graph convolutional network architecture, including a working condition feature recognition layer, a load trend evolution layer, and a power gap triggering layer. The operating condition feature identification layer: establishes a feature matrix containing typical load conditions, and determines the current load condition type by the matching degree between the real-time state parameter matrix and the feature matrix; wherein, the matching degree is calculated using cosine similarity. The load trend evolution layer: Based on the current load condition type, a multi-level prediction algorithm is adopted to adaptively match the preset combined model hyperparameters, and the first, second and third prediction sub-models are established respectively. The weighted outputs of each sub-model are spliced ​​together to generate a load trend prediction curve covering future time. The power gap triggering layer: calls the benchmark output of thermal power units to construct the basic output trajectory, calculates the time domain difference with the load trend prediction curve, and collects it into a gap sequence; extracts the amplitude and duration of the power gap sequence, sets a secondary warning threshold, and through a secondary discrimination mechanism, filters out cases where the amplitude is greater than the first warning threshold and triggers the first warning signal. At the same time, it calculates the product of amplitude and duration, filters out cases where the product result is greater than the second warning threshold, and triggers the second warning signal.

4. The method for load prediction and optimized scheduling of multi-energy storage thermal power units according to claim 3, characterized in that, The establishment of a feature matrix containing typical load conditions includes: Historical state parameter matrices covering periods of stable load, periods of sudden load change, and periods of sudden load drop are collected. The features of the historical state parameter matrix are denoised to separate low-frequency trend terms and high-frequency random terms, and the denoised feature matrix is ​​reconstructed. The correlation between the denoised wind speed, irradiance, temperature and ACG load command fluctuations is analyzed using Pearson correlation coefficient. Key parameter features are screened by the maximum mutual information coefficient.

5. The method for load prediction and optimized scheduling of multi-energy storage thermal power units according to claim 1, characterized in that, The decoupling consists of several components, including: The identified load disturbance sequence is iteratively decomposed using variational mode decomposition to obtain several intrinsic mode functions; Based on the intrinsic mode function, instantaneous amplitude and instantaneous frequency are extracted to construct a two-dimensional feature space, and a feature trajectory curve is formed by fitting. Based on the feature trajectory curve, the zero-crossing point and its direction of change are identified, and several monotonic fluctuation intervals are divided. Cluster analysis of monotonic fluctuation intervals identifies several components, including low-frequency, mid-frequency, and high-frequency components.

6. The method for load prediction and optimized scheduling of multi-energy storage thermal power units according to claim 1, characterized in that, The preset decision network includes a low-frequency reference branch, a mid-frequency buffer branch, and a high-frequency response branch. The low-frequency reference branch has a built-in first constraint, the mid-frequency buffer branch has a built-in second constraint, and the high-frequency response branch has a built-in third constraint.

7. The method for load prediction and optimized scheduling of multi-energy storage thermal power units according to claim 6, characterized in that, The constraints built into the decision network that are dynamically corrected based on the early warning signal include: A multidimensional constraint parameter space is constructed based on the low-frequency reference branch, the mid-frequency buffer branch, and the high-frequency response branch. The multidimensional constraint parameter space is orthogonally constructed by the first constraint, the second constraint, and the third constraint. At the same time, the boundary parameters of each constraint condition are extracted, including the boundary interval, the upper boundary limit, and the lower boundary limit. In conjunction with the early warning signal, the multidimensional constraint parameter space is dynamically reconstructed. Under the condition of responding to the first early warning signal, a first switching strategy is executed for the boundary parameters of the first constraint, and a second switching strategy is executed for the boundary parameters of the third constraint; under the condition of responding to the second early warning signal, a third switching strategy is executed for the boundary parameters of the second constraint, and a fourth switching strategy is executed for the boundary parameters of the third constraint. Latin hypercube sampling is used to perform hierarchical sampling in the multidimensional constraint parameter space to generate several uniformly distributed sample points; the sample points are combined to generate a correction vector; the correction vector contains the combination of boundary parameters corresponding to different constraint conditions.

8. The method for load prediction and optimal scheduling of multi-energy storage thermal power units according to claim 7, characterized in that, The first constraint includes fuel ramp rate and unit output amplitude; the second constraint includes temperature regulation dead zone and valve operating frequency; and the third constraint includes charging and discharging power and charging and discharging trigger boundary.

9. The method for load prediction and optimal scheduling of multi-energy storage thermal power units according to claim 1, characterized in that, The improved Dragonfly algorithm optimization iteration includes: The correction vector is used as the initial position vector of the dragonfly population in the Dragonfly Algorithm. Before each iteration, a state deduction operation is performed, including: updating the constraints in the decision network based on the correction vector; simultaneously, retrieving several currently obtained components and establishing the correlation between the updated constraints and the components; and extracting the first evaluation function and the second evaluation function based on the correlation. The first evaluation function is as follows: extract the amplitude of the low-frequency component and calculate the ratio with the corresponding first constraint, and mark it as the first index; extract the amplitude of the mid-frequency component and calculate the difference with the corresponding second constraint, and mark it as the second index; construct the first evaluation function based on the first index and the second index. The second evaluation function is as follows: extract the amplitude of the high-frequency component and calculate the difference with the corresponding third constraint, and mark it as the third index; extract the frequency of change of the mid-frequency component and calculate the deviation from the second constraint, and mark it as the fourth index; construct the second evaluation function based on the third index and the fourth index. In each iteration, the position vector is updated by maximizing the first evaluation function and minimizing the second evaluation function until the maximum number of iterations is satisfied.

10. A load forecasting and optimized scheduling system for multi-energy storage thermal power units, characterized in that, The system includes: The data acquisition module collects operating parameters of the thermal power unit and external environmental parameters in real time, and constructs a real-time status parameter matrix; The load forecasting module constructs a load forecasting model based on the historical state parameter matrix, imports the real-time state parameter matrix into the load forecasting model, outputs a load trend forecasting curve, and triggers an early warning signal through a two-level discrimination mechanism. The optimized scheduling module identifies load disturbance values ​​based on the load trend prediction curve, obtains the load disturbance sequence, and decouples it into several components. Several vectors are input into a preset decision network, and the constraints built into the decision network are dynamically corrected in combination with the early warning signal. An improved dragonfly algorithm is introduced to iterate and generate optimized scheduling instructions.

Images