Flywheel-battery hybrid energy storage system and hierarchical coordination control method thereof

Abstract

The application discloses a flywheel-battery composite energy storage system and a layered coordination control method thereof, and belongs to the technical field of new energy power generation and energy storage. The system comprises an energy management layer, a coordination distribution layer and a device control layer. The energy management layer generates total power instructions based on ultra-short-term prediction and scheduling instructions. The coordination distribution layer uses power fluctuation spectrum analysis to dynamically distribute high-frequency components to the flywheel to cope with second / sub-second impact, and to distribute low-frequency components to the battery to realize energy time shift, and to realize power non-inductive switching when the flywheel reaches the limit. The device control layer respectively embeds a temperature adaptive algorithm for the mechanical bearing of the flywheel and the electrochemical characteristics of the battery, and optimizes the operating parameters under extreme temperature difference to ensure the safety and service life of the device. The application effectively solves the problem of service life attenuation caused by high power impact of traditional battery energy storage in desert scenes, and significantly improves the smoothness of power station output and the stability of power grid frequency.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation and energy storage technology, and in particular to a flywheel-battery composite energy storage system for wind-solar-storage power stations in extreme environments such as deserts and Gobi, and its hierarchical coordinated control method. Background Technology

[0002] With the acceleration of the global energy transition, the large-scale development of integrated wind-solar-storage power plants in desert and Gobi regions has become an important way to achieve the "dual carbon" goal. However, such scenarios face multiple technical challenges, mainly in the following aspects:

[0003] First, the output of photovoltaic power in desert areas is affected by sandstorms and shading, resulting in intermittent drops on the order of seconds. The output of wind power is affected by gusts of wind, resulting in random fluctuations on the order of sub-seconds. The combination of the two forms a broadband power surge.

[0004] Secondly, traditional single-battery energy storage systems need to frequently cope with high-rate charge and discharge shocks, which leads to repeated rupture and reconstruction of the SEI film inside the cell, continuous consumption of active lithium, and a cycle life that is reduced to less than 60% of the nominal value.

[0005] Third, the diurnal temperature range in desert areas can reach more than 40°C, and the change rate of battery internal resistance exceeds 200% under extreme temperatures. Existing fixed-parameter control strategies cannot adapt to this, resulting in the coexistence of high-temperature thermal runaway and low-temperature lithium precipitation risks.

[0006] Fourth, most existing hybrid energy storage systems adopt a static stratification strategy based on the state of charge (SOC), which fails to achieve functional decoupling between power-type and energy-type energy storage, resulting in an inoptimal match between the flywheel's high-frequency regulation capacity and the battery's energy time-shifting capability.

[0007] In other words, traditional battery energy storage will experience a decline in lifespan due to the high power surges it faces in desert environments. Summary of the Invention

[0008] Purpose of the invention:

[0009] To address the issue of lifespan degradation under high-power surges in single-cell energy storage, this invention provides a flywheel-battery hybrid energy storage system and its hierarchical coordinated control method.

[0010] Technical solution:

[0011] In a first aspect, the present invention provides a flywheel-battery hybrid energy storage system, comprising:

[0012] The energy management layer configures the power reference commands used to generate the flywheel and battery.

[0013] The coordination and allocation layer is configured to perform power spectrum splitting on the power reference command, allocating high-frequency components as flywheel power commands and low-frequency components as battery power commands; and defines flywheel regulation capability based on the flywheel speed operating boundary, smoothly transferring power components exceeding the flywheel regulation capability to the battery at a preset slope; and

[0014] The device control layer is configured to perform temperature adaptation control, including: temperature adaptation control of the flywheel to cope with high-frequency power fluctuations; and temperature adaptation control of the battery to achieve energy time shift and power smoothing.

[0015] The above solution leverages the physical advantages of flywheels in coping with second / sub-second power surges and the chemical advantages of batteries in achieving minute-hour energy shifts through dynamic high-frequency-low-frequency power allocation, thus solving the problem of lifespan degradation under high-power surges in single-battery energy storage.

[0016] Furthermore, an extreme temperature adaptive control system was established for the extreme temperature difference environment in the desert. Temperature adaptive algorithms were embedded in the flywheel bearing system and the battery electrochemical system to achieve intelligent management of high temperature derating protection and low temperature preheating start-up, ensuring the safety and lifespan of the equipment in all climates.

[0017] Furthermore, through closed-loop speed verification and smooth transfer algorithms, seamless switching of power commands is achieved. When the flywheel adjustment capability reaches its limit, the excess power component is automatically transferred to the battery at a preset slope, eliminating the risk of power sudden changes and ensuring the continuity of power adjustment.

[0018] In a second aspect, the present invention provides a hierarchical coordinated control method based on the flywheel-battery composite energy storage system described in the first aspect, comprising the following steps:

[0019] Construct a power prediction model;

[0020] A standard quadratic programming form is established, which includes an objective function and constraints. The objective function adopts a weighted summation of three objectives and minimizes each performance index. At the same time, multi-dimensional constraints are established, including power balance constraints, time-varying power limit constraints, state boundary constraints, and power change rate constraints. Among them, the upper and lower limits of the power in the time-varying power limit constraints are dynamically adjusted according to real-time temperature and state.

[0021] The optimal control sequence is obtained by using a sequential quadratic programming algorithm. The first element of the optimal control sequence is used as the control variable at the current moment, which includes the battery power command and the flywheel power command.

[0022] The battery power command and flywheel power command are sent to the coordination and allocation layer; at the same time, the measured changes in the state of charge and state of energy of the energy storage system are collected, and the model prediction error is calculated; if the error exceeds the set threshold, the prediction model parameters or initial state estimate are adjusted through state feedback correction in the next control cycle.

[0023] Rolling time-domain update and iterative optimization.

[0024] Furthermore, the step of constructing the power prediction model includes:

[0025] Multi-source feature extraction and input vector construction;

[0026] A two-layer LSTM is used for temporal feature extraction; and a Dropout regularization mechanism is introduced after the output layer of the second LSTM; the feature vector after Dropout is linearly transformed through a fully connected layer and mapped to the prediction output space.

[0027] The power prediction model is trained and optimized based on a loss function that incorporates a gust mutation penalty term;

[0028] Backpropagation and parameter optimization;

[0029] Output of ultra-short-term prediction sequences.

[0030] Beneficial effects:

[0031] The significant advantages of this invention compared to existing technologies are:

[0032] 1. This invention achieves functional decoupling and unified scheduling of power-type energy storage and energy-type energy storage through the deep integration of flywheel energy storage and battery and the hybrid access topology of DC bus. It effectively leverages the complementary advantages of flywheel in dealing with high-frequency impacts at the second / sub-second level and battery in achieving energy time shifts at the minute to hour level, significantly reducing the lifespan degradation caused by frequent high-rate charging and discharging of a single battery and extending the entire life cycle of the energy storage system.

[0033] 2. Based on power fluctuation spectrum analysis and a three-layer collaborative hierarchical intelligent control architecture, the present invention can accurately identify the high-frequency and low-frequency components of wind and solar power output and dynamically allocate them to the flywheel and battery. It realizes seamless and smooth switching of power commands when the flywheel is at its adjustment limit, effectively improving the smoothness of power plant output and the stability of grid frequency, and avoiding equipment overload and regulation interruption caused by power surges.

[0034] 3. This invention combines ultra-short-term power prediction, model predictive control rolling optimization, and real-time feedback compensation for extreme temperatures. The system can dynamically adapt to the dual disturbances of strong random wind and solar power output and large temperature differences between day and night in desert areas, ensuring the safe and stable operation of flywheel bearings and battery cells in extreme environments ranging from -30℃ to +70℃. This significantly enhances the adaptability and reliability of the energy storage system under harsh climatic conditions.

[0035] 4. This invention achieves plug-and-play upgrades to existing wind and solar power station equipment through a topology design of single-point DC bus access and unified inverter grid connection. Power regulation can be completed without modifying wind turbines and photovoltaic inverters, effectively reducing the engineering difficulty and modification cost of system deployment, and improving the integration and compatibility of energy storage systems with existing power stations. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the structure and flow of the flywheel-battery hybrid energy storage system of the present invention;

[0037] Figure 2 This is a schematic diagram of the steps of the hierarchical coordination control method of the present invention;

[0038] Figure 3 This is one of the schematic diagrams illustrating the effect gain of the present invention compared to the conventional system;

[0039] Figure 4 The second diagram illustrates the effect gain of the present invention compared to the conventional system. Detailed Implementation

[0040] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and not intended to limit the scope of the invention. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the embodiments of the present invention, and not all structures.

[0041] In the following description, specific details such as target system architecture and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0042] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0043] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0044] Furthermore, in the description of this application and the appended claims, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0045] References to "one embodiment" or "some embodiments" in this specification mean that one or more embodiments of this application include the target features, structures, or characteristics described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0046] like Figure 1 The flywheel-battery hybrid energy storage system shown is used for wind-solar-storage integrated power stations in desert environments. The energy storage system includes a top-level energy management layer, a middle-level coordination and distribution layer, a bottom-level equipment control layer, and a power electronic conversion and grid connection system.

[0047] The top-level energy management layer collects real-time power output data and extreme temperature data of energy storage power stations, and combines wind and solar ultra-short-term forecasts with grid status feedback. With the optimization goal of smoothing the total power output of the power station and participating in grid frequency regulation, it continuously optimizes and generates the power reference command for the flywheel and battery within the next 15 minutes. The middle-level coordination and allocation layer then performs power spectrum splitting on the power reference command, allocating the high-frequency component as the flywheel power command and the low-frequency component as the battery power command.

[0048] Furthermore, in response to the extreme temperature differences in the desert environment, the underlying equipment control layer will implement two aspects of control:

[0049] Temperature-adaptive control of the flywheel is implemented to cope with high-frequency power fluctuations;

[0050] Temperature-adaptive control of the battery is implemented to achieve energy time shift and power smoothing.

[0051] Furthermore, the power electronic conversion and grid connection system aggregates the output power of the flywheel and battery to the DC bus, which is then converted by a high-power inverter and connected to the grid or supplied to the load. The system also feeds back the equipment status and grid system status to the top-level energy management layer.

[0052] In one specific implementation, the central coordination controller in the coordination and allocation layer receives the flywheel / battery power reference command issued by the top-level energy management layer, and simultaneously receives data reported by the bottom-level flywheel / battery controller, namely:

[0053] Flywheel current speed, available charging and discharging power, bearing temperature limit;

[0054] Real-time operating boundaries for battery current state of charge (SOC), charge / discharge rate limits, and temperature health status.

[0055] The synchronous reception of the aforementioned operational data is intended to combine the initial power allocation results obtained from spectrum splitting with the real-time executable boundaries of the equipment. Specifically, the central coordinating controller calculates the remaining charging and discharging capabilities of the flywheel and battery online based on the synchronously received data, and dynamically adjusts the upper and lower power limits. When the flywheel speed, bearing temperature, or available power approaches the boundary, or when the battery SOC, charge / discharge rate, and temperature limits tighten, power limiting and reallocation are performed in a timely manner. This avoids unexecutable commands, flywheel overspeed or underspeed, and battery overcharging or over-discharging, thus forming a closed-loop control process of command allocation, status feedback, and boundary updates.

[0056] To perform power spectrum decomposition on the power reference command, in this embodiment, the central coordination controller has a built-in spectrum analysis module. This module performs a Fast Fourier Transform on the power reference command to decompose the command into high-frequency and low-frequency components, wherein:

[0057] High-frequency components are power components with a variation period of less than 1 minute, large amplitude fluctuations, and short durations, such as power surges caused by sudden changes in wind power gusts and photovoltaic cloud shading, and sub-second frequency disturbances in the power grid, all of which are allocated to the flywheel;

[0058] Low-frequency components are power components with a change period of ≥1 minute, gradual change, and long duration, such as power plant output plan deviations and energy time shift demands, all of which are allocated to lithium iron phosphate batteries.

[0059] Furthermore, after allocating power commands based on high-frequency and low-frequency components, the central coordination controller verifies in real time whether the allocated power commands exceed the equipment's safety boundaries. The core of this verification is a closed-loop check of the flywheel speed. This real-time verification ensures that the spectrum allocation results are executable within the equipment's physical constraints. By monitoring the flywheel speed and remaining adjustable capacity, the risk of flywheel saturation (i.e., operating boundaries) can be identified in advance. Before reaching the operating boundaries, the excess portion is smoothly transferred to the battery at a defined slope, thereby preventing flywheel overspeed, underspeed, or power saturation. Simultaneously, it suppresses command steps and DC bus voltage fluctuations caused by power switching, ensuring the continuity and safety of the hybrid energy storage system's regulation process.

[0060] Specifically, when the flywheel speed approaches the upper limit of 95% of the rated speed or the lower limit of 20% of the rated speed, or when the adjustable capacity is insufficient, a smooth transfer algorithm is triggered to smoothly transfer the power component that exceeds the flywheel's adjustment capability to the battery at a preset slope, avoiding sudden power changes. When the flywheel speed returns to the safe operating range, the transferred power component is smoothly withdrawn, and the optimal allocation logic of high-frequency flywheel and low-frequency battery is restored.

[0061] Considering that the above implementation uses a DC bus coupled topology, preferably, the battery-side execution unit is a bidirectional DC / DC converter, and a unified grid-connected inverter is set up on the grid-connected side. The final flywheel real-time power command is sent to the flywheel local controller through a low-latency communication bus, which controls the flywheel-side bidirectional AC / DC converter; the battery real-time power command is sent to the battery bidirectional DC / DC controller through a low-latency communication bus.

[0062] It should be understood that the core positioning of the underlying device control layer in this embodiment is precise execution of power commands and adaptive protection against extreme temperatures. Specifically, it is divided into two parallel units: the flywheel energy storage control branch and the battery energy storage control branch.

[0063] The specific implementation method of the flywheel energy storage control branch is as follows:

[0064] The flywheel local controller receives real-time power commands from the middle layer and simultaneously collects real-time flywheel speed, rotor position, motor winding temperature, air / magnetic bearing temperature, vibration amplitude, vacuum level, and DC bus voltage and current operating data at high speed. Through the flywheel local controller's built-in optimal operating point tracking algorithm adapted to bearing temperature characteristics, it is specifically optimized for extreme desert temperature scenarios. Specifically, in extreme summer high-temperature environments, as bearing temperature rises, load-bearing capacity and maximum speed decrease, the algorithm automatically lowers the flywheel charging and discharging power limits and narrows the safe speed range to prevent bearing overheating and damage. In extreme winter low-temperature environments, as bearing lubrication characteristics change, the algorithm first executes low-speed preheating logic, releasing rated power only after the bearing temperature rises to a reasonable range, ensuring mechanical safety and service life.

[0065] As an exemplary implementation, the measured bearing temperature is used as the primary criterion, with ambient temperature as a secondary criterion. When the ambient temperature is not lower than 45°C or the bearing temperature is not lower than 70°C, corresponding to extreme high-temperature environments in summer, the system enters the high-temperature derating operation zone. As the temperature rises, the controller gradually lowers the upper limit of charging and discharging power and narrows the safe range of flywheel speed. Specific thresholds can be calibrated according to the type of air bearing or magnetic bearing, material characteristics, insulation class, and the allowable temperature specified by the manufacturer, and are not limited to the values ​​in the example above.

[0066] When the ambient temperature is not higher than -20℃ or the bearing temperature is not higher than 0℃, corresponding to the extreme low temperature environment in winter, it enters the low temperature start-up zone, first operating at low speed and low power to complete preheating; after the bearing temperature rises to above 5℃ and remains stable, the rated power capacity is gradually released according to the temperature change. The specific temperature range can be calibrated according to the bearing type and the low temperature performance of the materials.

[0067] Furthermore, the flywheel local controller employs a field-oriented vector control algorithm to control the switching state of the bidirectional AC / DC converter, driving a high-speed permanent magnet motor to precisely track power commands. When the energy storage system is charging, the motor operates in motor mode, accelerating the flywheel rotor and converting electrical energy into rotor kinetic energy for storage, absorbing excess power from wind and solar power. When the energy storage system is discharging, the motor operates in generator mode, the flywheel rotor decelerates, converting kinetic energy into electrical energy, which is then rectified by the converter and fed into the DC bus of the power electronic conversion and grid-connected system, smoothing out power surges and frequency fluctuations, and supplementing the power output gap caused by wind and solar power.

[0068] At the same time, the flywheel's real-time operating status, available power limits, and fault information are reported to the top-level energy management layer in real time, forming a closed-loop control.

[0069] The specific implementation method of the battery energy storage control branch is as follows:

[0070] The bidirectional DC / DC controller receives real-time battery power commands from the middle layer, while the Battery Management System (BMS) collects real-time data on the voltage and temperature of individual cells, the total voltage of the battery cluster, the charging and discharging current, the state of charge, the state of health, the insulation resistance, and the battery cluster status.

[0071] The BMS works in conjunction with the DC / DC controller and incorporates a battery electrothermal coupling model. It enables adaptive battery life compensation for extreme high and low temperature scenarios, such as large day-night temperature differences in desert environments. Specifically, in summer, extreme high temperatures can cause cell temperatures to rise, accelerating the aging of the solid electrolyte interface film and reducing cycle life. The algorithm automatically lowers the battery charge / discharge rate limits and narrows the charge / discharge voltage window to prevent high-rate charging and discharging from exacerbating temperature rises, prioritizing battery health. In winter, extreme low temperatures can lead to decreased cell activity, and high-rate charging and discharging can cause lithium deposition. The algorithm first activates the thermal management system to preheat the cells, while simultaneously dynamically lowering the charge / discharge rate and adjusting the charging cutoff voltage to avoid irreversible damage at low temperatures, significantly extending battery life in desert environments.

[0072] As an exemplary implementation, it is recommended to use cell temperature as the criterion for high-temperature control of the battery. When the cell temperature is not lower than 45°C, corresponding to extreme high-temperature environments in summer, the battery enters the high-temperature derating zone, dynamically reducing the battery charge / discharge rate and narrowing the charge / discharge voltage window. When the cell temperature reaches the high-temperature protection threshold specified by the manufacturer, the corresponding charging or discharging operation is stopped. The specific threshold should be calibrated according to the technical specifications of the lithium iron phosphate cells used.

[0073] When the cell temperature is not higher than 0℃, corresponding to the extreme low temperature environment in winter, it enters the low temperature control zone, limits the battery charging rate, and activates the thermal management system to preheat the cell; after the cell temperature rises to above 5℃, the charging capacity is gradually restored, and the battery discharge power is also dynamically reduced according to the temperature. The above 0℃ and 5℃ are exemplary thresholds, and the actual values ​​can be calibrated according to the allowable charging and discharging temperature range of the cell.

[0074] Furthermore, the bidirectional DC / DC controller adjusts the duty cycle of the switching transistors using pulse width modulation (PWM) based on power commands, precisely controlling the charging and discharging current and power of the battery. When the energy storage system is charging, it absorbs electrical energy from the DC bus in the power electronic conversion and grid-connected system, matching the battery cluster voltage to charge the battery, storing excess wind and solar energy, and realizing energy time shifting. When the energy storage system is discharging, the battery discharges, and after DC / DC conversion, it is fed into the DC bus in the power electronic conversion and grid-connected system, supplementing the low-frequency gap in wind and solar power output and smoothing the power output curve of the power station from minutes to hours.

[0075] Meanwhile, the BMS reports the battery's real-time SOC / SOH, available charge / discharge power limits, temperature status, and fault information to the top-level energy management layer, providing constraints for upper-level optimization and power allocation.

[0076] Furthermore, the power electronic conversion and grid-connected system connects to the same DC bus in parallel through the output sides of the flywheel bidirectional AC / DC converter and the battery bidirectional DC / DC converter, realizing single-point access and unified scheduling of the flywheel and battery. There is no need to modify the wind turbine or photovoltaic inverter, and all power regulation is completed on the DC bus side.

[0077] Specifically, the grid-connected inverter maintains a stable DC bus voltage through DC voltage closed-loop control, providing a reliable DC voltage platform for the flywheel and battery converters. The inverter controller synchronizes with the grid through a phase-locked loop (PLL) and employs a grid voltage-oriented vector control algorithm to precisely control active and reactive power output, converting the DC bus power into AC power that conforms to grid standards, which is then fed into the power plant's AC combiner system and ultimately connected to the grid or supplied to local loads.

[0078] Meanwhile, the power electronic conversion and grid connection system is also used to respond to commands from the top-level Energy Management System (EMS), participate in grid frequency and voltage regulation, provide reactive power support, possess low / high voltage ride-through capabilities, adapt to grid fault scenarios, and ensure the grid connection stability of desert power plants. Furthermore, it collects grid voltage, frequency, harmonics, and other status data in real time and reports them to the top-level EMS, providing grid-side constraints for energy management optimization.

[0079] In another embodiment, a hierarchical coordinated control method is implemented based on the aforementioned flywheel-battery composite energy storage system. Specifically, the redundantly configured energy storage EMS industrial controller in the energy management layer collects meteorological and extreme temperature data through a wide-temperature total radiation meter, anemometer, ambient temperature sensor, and wind and solar equipment temperature sensor; it collects gate metering data, real-time wind and solar power output in the desert, and grid frequency / voltage status data through the power station gate metering CT / PT, wind and solar power output acquisition module, and grid frequency / voltage acquisition device; and it collects the state of energy (SOE) of the bottom flywheel, the SOC of the battery, and the operating status data of both, including temperature, health, and available power limits, through the energy storage status acquisition unit.

[0080] To address the sampling jump problem caused by strong electromagnetic interference in the desert, this implementation uses a Kalman filter combined with a median filter algorithm to remove outliers, and uses GPS timestamp alignment to ensure that the synchronization accuracy of multi-source data is better than 10ms.

[0081] like Figure 2 The hierarchical coordination and control method shown includes:

[0082] The specific steps for constructing a power prediction model for an EMS host running a Long Short-Term Memory (LSTM) network are as follows:

[0083] Step 1: Multi-source feature extraction and input vector construction, specifically:

[0084] The input consists of multi-dimensional data collected from the desert power plant site, which is used to construct a time-series input feature vector, including:

[0085] Real-time meteorological sequences of irradiance G, wind speed V, and ambient temperature Tamb;

[0086] Historical output patterns of photovoltaic / wind power under similar weather conditions in the previous 24 hours;

[0087] Extreme scenario indicators for sandstorm shading probability (Pdust) and gust change index (GFI).

[0088] The above features are concatenated along the time dimension to form the input feature vector at the current time t. Where n is the feature dimension, It is the set of real numbers.

[0089] Step 2: Temporal feature extraction using a two-layer LSTM, specifically:

[0090] Input vector The data is fed into a two-layer LSTM for deep temporal feature extraction. The first layer, an LSTM with 128 hidden units, receives the original input and updates the cell state through the synergistic action of the forget gate, input gate, and output gate, capturing short-term fluctuation features. Its output serves as the input to the second layer, an LSTM with 64 hidden units, to further extract medium- and long-term temporal dependencies.

[0091] It should be noted that the state update of an LSTM cell follows the following mechanism:

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] ;

[0097] ;

[0098] In the formula, t is the index of the current time step; Let be the input feature vector at time t; This is the hidden state from the previous moment; This indicates vector concatenation; , and These are the activation vectors for the forget gate, input gate, and output gate, respectively. and These represent the cell states at the current and previous moments, respectively. Candidate cell state; Currently hidden; , , and These are the weight matrices for the corresponding gate structure or candidate state; , , and For the corresponding bias vector; Use the Sigmoid activation function; It is the hyperbolic tangent function; This indicates element-wise multiplication; the subscripts f, i, o, and C represent the forget gate, input gate, output gate, and candidate cell state, respectively.

[0099] By employing a two-layer stacking mechanism, the model effectively captures the complex temporal patterns of wind and solar power output in desert environments. Specifically, the first-layer LSTM uses a gating mechanism to selectively memorize irradiance, wind speed, and real-time power output sequences, focusing on extracting second- and minute-level local abrupt changes caused by gusts and dust shading. Its cell state provides long-path transmission of historical information, mitigating gradient vanishing during long-sequence training. The second-layer LSTM, based on the short-term features extracted by the first layer, further integrates power output patterns under similar weather conditions from the previous 24 hours, as well as extreme scenario markers such as dust shading probability (Pdust) and GFI, to extract continuous changing trends and cross-temporal dependencies. This forms a hierarchical temporal representation of local abrupt changes—medium- to long-term evolutionary trends—and extreme event priors, thereby improving the accuracy and robustness of wind and solar power output prediction under complex perturbation scenarios in deserts.

[0100] Step 3: Dropout regularization is applied to force the model to learn more robust feature representations, enhancing its generalization ability to scenarios with strong electromagnetic interference in deserts and sudden weather changes. Specifically:

[0101] To prevent overfitting during training, a Dropout regularization mechanism is introduced after the output layer of the second LSTM layer. For example, some neuron connections are randomly dropped with probability p=0.2, thus generating a mask vector that follows a Bernoulli distribution.

[0102] ;

[0103] In the formula, m is the Bernoulli (Dropout) mask vector, which follows a Bernoulli distribution with parameter 1−p; p is the probability of dropping a neuron, which is 0.2 in this embodiment; 1−p is the probability of retaining a neuron; the elements in the mask vector are either 0 or 1, where 0 indicates that the corresponding hidden unit is temporarily masked in this training iteration, and 1 indicates that it is retained; the Dropout mechanism is only enabled during the model training phase.

[0104] Then, perform element-wise multiplication between the mask vector and the hidden layer output:

[0105] ;

[0106] In the formula, This represents the hidden feature vector output by the second LSTM layer at time t. is the feature vector after Dropout processing; m is the Bernoulli mask vector; This represents element-wise multiplication. If a reverse Dropout implementation is used, scaling can be performed during the training phase according to the retention probability 1−p to maintain consistency in the mathematical expectation of the output between the training and inference phases.

[0107] Step 4: Generating the fully connected layer and prediction output, specifically:

[0108] The feature vectors processed by Dropout undergo a linear transformation through a fully connected (Dense Layer) to map them to the prediction output space. Specifically, two parallel output heads are used, each containing 15 neurons, to generate photovoltaic power output prediction sequences and wind power output prediction sequences, respectively:

[0109] ;

[0110] In the formula, s is the energy type index; PV represents photovoltaic; wind represents wind power; Let be the predicted output value of energy type s at time t for the k-th step in the future; k=1,…,15; and These are the weight matrix and bias vector of the output head corresponding to energy type s, respectively; The time-series features are after Dropout processing. Therefore, the photovoltaic and wind power prediction sequences share the time-series features extracted by LSTM, but are mapped separately using different output parameters.

[0111] Step 5: Calculate the improved loss function, specifically:

[0112] To address the drastic power output fluctuations caused by sudden wind speed changes and sandstorm shading in desert regions, the model employs an improved composite loss function for training and optimization. This loss function, based on the traditional mean squared error (MSE), introduces a penalty term for sudden wind gusts to enhance the model's ability to recognize extreme scenarios. The specific expression is as follows:

[0113] ;

[0114] In the formula, The loss function is a composite function; N is the number of training samples; i is the index of the training sample. and These are the predicted output and actual output of the i-th sample, respectively; This is the penalty coefficient for sudden gust changes. ; Let i be the gust change indicator function corresponding to the i-th sample; This represents the actual output value. The function indicating sudden gusts is defined as follows:

[0115] ;

[0116] When the rate of change of wind speed exceeds 5 m / s / min (m / s / min represents the change in wind speed per minute, that is, the wind speed changes by 5 m / s in 1 minute, and the unit can be written as...), When the threshold of ) is reached, Setting it to 1 means that even if the prediction error is small, it will still be compensated by the penalty term. By increasing the loss value, the model is forced to focus on the prediction accuracy at the moment of sudden gust changes during training, thereby reducing the overall prediction bias by 15%-20%.

[0117] Step 6, backpropagation and parameter optimization, specifically:

[0118] Based on the loss function value calculated in step 5 The model calculates the gradients of the parameters in each layer using the Backpropagation Through Time (BPTT) algorithm and updates the network weights using the Adaptive Moment Estimation (Adam) optimizer, the expression of which is:

[0119] ;

[0120] In the formula, The network parameters were as before the update. For the updated network parameters; The learning rate; This is the first-order moment estimate after bias correction in the t-th optimization iteration; This is a second-order moment estimate after bias correction; A tiny positive number set to prevent the denominator from being zero; t is the number of optimization iterations.

[0121] Through iterative optimization, the loss function of the model on the validation set converges to its minimum value.

[0122] Step 7: Output the ultra-short-term prediction sequence, specifically:

[0123] After training, the model enters online prediction mode, inputting measured real-time meteorological data, historical power output curves, and extreme scenario markers into the trained network. After forward computation, it directly outputs a photovoltaic power output prediction sequence for the next 15 minutes with a time resolution of 1 minute. Wind power output prediction sequence This prediction sequence serves as input to the top-level energy management layer's MPC optimization, generating power reference commands for the energy storage system and achieving closed-loop coordination between prediction and scheduling.

[0124] Furthermore, the redundant EMS controller performs rolling optimization calculations based on model predictive control (MPC) in 1-minute increments. The specific steps are as follows:

[0125] Step I: Initialize MPC control parameters and prediction time domain settings, specifically:

[0126] The time frame and optimization dimensions of model predictive control, and the system settings for the prediction time domain. Control the time domain for the next 15 minutes. The sampling period corresponds to the controllable interval of the first 5 minutes. Simultaneously, initialize the multi-target weight coefficients, including the tracking weight. (Usually taken as 0.5), frequency modulation weight (Usually taken as 0.3), lifetime weight (Typically taken as 0.2), and the device's rated parameters include the battery's rated capacity. Flywheel rated capacity Battery efficiency Flywheel efficiency These parameters constitute the basic constraint boundary for MPC optimization.

[0127] Step II, Real-time Status Measurement and Parameter Identification, specifically:

[0128] At the start of each control cycle, t, the EMS controller acquires the system state vector in real time through the sensor network, providing accurate initial conditions for the prediction model. The measured quantities include: the battery's current state of charge (SOC(t), the flywheel's current state of energy storage (SOE(t)), the cell temperature (Tcell(t), the flywheel bearing temperature (Tbearing(t)), the flywheel's current rotational speed (n(t)), and the photovoltaic output prediction value for the next 15 minutes from the LSTM ultra-short-term prediction model. Wind power output forecast (k=1,2,…,15). Simultaneously, the current grid frequency deviation Δf(t) and the dispatch plan curve Pplan(t) are collected as benchmarks for optimization.

[0129] Step III: Establish a discrete-time prediction model, specifically:

[0130] Based on the initial state obtained in step II, a discrete state-space equation describing the future dynamic evolution of the energy storage system is constructed. The discrete-time state prediction model inside the MPC is used to predict the energy storage state evolution process, which is specifically composed of the battery SOC state transition equation and the flywheel SOE state transition equation.

[0131] The battery SOC dynamic equation describes the change in battery state of charge with charge and discharge power:

[0132] ;

[0133] The flywheel SOE dynamic equations describe the evolution of the flywheel's energy storage state as a function of power command:

[0134] ;

[0135] In the formula, and These represent the battery state of charge and flywheel energy state predicted at the current time t based on known information at step k, respectively; t is the current control time; k is the prediction step size index; the vertical bar "|t" indicates that the known information at time t is used as a condition; The sampling period; and These are the battery power command and flywheel power command to be optimized at the k-th step in the prediction time domain, respectively. Positive values ​​indicate discharging and negative values ​​indicate charging. and These are the energy conversion efficiencies of the battery and flywheel, respectively. and These are the rated available energy values ​​for the battery and flywheel, respectively. The subscripts bat and fly represent the battery and flywheel, respectively.

[0136] These two equations together form the state transition matrix of MPC, supporting the prediction of the state trajectory in the next 15 minutes.

[0137] Step IV: Construct a constrained multi-objective optimization problem, specifically:

[0138] The power plant's operational requirements are transformed into a mathematical optimization problem, establishing a standard quadratic programming form that includes an objective function and constraints. The objective function adopts a weighted summation of three objectives, minimizing the following performance indices:

[0139] ;

[0140] The first term is the tracking error term, which minimizes the deviation between the grid-connected power and the scheduling plan, where:

[0141] ;

[0142] The second term is the frequency support term, which minimizes the sum of squares of the grid frequency deviation and dynamically increases when the frequency deviation exceeds the dead zone. ;

[0143] The third item is the lifespan decay item. Characterizes the energy storage aging cost coupled with temperature and rate.

[0144] Simultaneously, establish multi-dimensional constraints to ensure equipment safety:

[0145] Power balance constraints (equality constraints):

[0146] ;

[0147] Time-varying power limit constraints (inequality constraints):

[0148] ;

[0149] ;

[0150] The power upper and lower limits are dynamically adjusted based on real-time temperature and status; the limits are automatically lowered at high temperatures and raised at low temperatures. It also includes state boundary constraints.

[0151] ;

[0152] And power change rate constraints:

[0153] .

[0154] In the above formula, J represents the overall optimization objective; k is the prediction time domain length; k is the prediction step size; , and The weights are respectively for the grid-connected power tracking term, frequency support term, and lifetime degradation term; To predict grid-connected power; For scheduling planned power; To predict frequency deviation; This represents the equivalent lifespan degradation cost of the battery and flywheel. In the power balance equation, , and These are the predicted power output of photovoltaic power, the predicted power output of wind power, and the load power, respectively. and These represent the power of the battery and the flywheel, respectively. After transforming to the standard optimization form, the state variables, control variables, and disturbance variables should be mapped to the state vector x, control vector u, and external input vector d, respectively, and the correspondence between each variable and the original physical quantity should be explained.

[0155] Step V: Solve using the Sequential Quadratic Programming (SQP) algorithm, specifically:

[0156] The sequential quadratic programming algorithm is used to solve the nonlinear constraint optimization problem established in step four. At the current iteration point... We perform a second-order Taylor expansion on the objective function and first-order linearization on the constraints to construct a quadratic programming (QP) subproblem:

[0157] ;

[0158] In the above formula, j is the number of SQP iterations; This is the control sequence corresponding to the j-th iteration; This is the control increment for this QP subproblem; The Hessian matrix of the objective function with respect to the control variables or its BFGS approximation matrix; The gradient of the objective function; and These are the coefficient matrix and constant vector of the linearization equality constraints, respectively; and These are the coefficient matrix and upper bound vector of the linearized inequality constraint, respectively; the superscript T indicates matrix transpose; st indicates "satisfies the following constraints".

[0159] The search direction can be obtained by solving the QP subproblem using the interior point method or the effective set method. The step size is determined by the meridian search. Post-update control sequence When the relative rate of change of the objective function The iteration may terminate if the number of iterations exceeds 50, thus obtaining the optimal control sequence for the next 15 minutes. and .

[0160] In the above formula, The step size is obtained by the meridian search in the j-th iteration; The updated control sequence; and These are the objective function values ​​for two consecutive iterations; This is the termination threshold for the relative rate of change of the objective function. and These represent the optimal battery power and optimal flywheel power for the k-th future step, respectively, with the asterisk "*" indicating the optimal solution.

[0161] Step VI, Control Command Implementation and Feedback Correction, specifically:

[0162] Only the first element of the optimal control sequence (the control quantity at the current time t) is implemented, i.e., the battery power command is issued. and flywheel power command The data is then distributed to the middle layer for coordination. Simultaneously, measured changes in SOC and SOE of the energy storage system are collected and compared with the predicted values ​​from step III to calculate the model prediction error.

[0163] ;

[0164] In the formula , This represents the SOC prediction error. and These represent the measured and predicted values ​​of SOC at the next time step, respectively; the SOE prediction error is defined in the same way. The subscripts actual and predicted represent the measured and predicted quantities, respectively, and bat and fly represent the battery and flywheel, respectively.

[0165] If the error exceeds the set threshold, the prediction model parameters or initial state estimate will be adjusted through state feedback correction in the next control cycle to compensate for the effects of model mismatch and external disturbances, thus forming a feedback correction mechanism.

[0166] Step VII, Rolling Time-Domain Update and Iterative Optimization, specifically:

[0167] The control time domain is rolled forward by one sampling period, t→t+1, and execution starts again from step II. This allows MPC to continuously use the latest measurement information to correct future predictions, effectively cope with the strong randomness and model uncertainty of desert landscape output, and ensure that the control strategy is always generated based on the current optimal information, achieving a dynamic balance between 15-minute look-ahead optimization and 1-minute real-time control.

[0168] By iteratively executing steps I to VII above, the MPC rolling optimization algorithm can achieve multi-objective collaborative optimization of power plant output smoothing, grid frequency support, and energy storage lifetime protection, while meeting extreme temperature constraints and equipment safety limits.

[0169] It should be understood that the 15-minute prediction time domain and 1-minute sampling period are merely preferred embodiments and not the sole limitation on the scope of protection of this invention. The prediction time domain and sampling period can be adjusted according to the characteristics of wind and solar power fluctuations, grid dispatch cycles, prediction accuracy requirements, controller computing power, and communication refresh frequency. The prediction time domain meets the following requirements. ,in To predict the number of steps, The sampling period is 15 min for prediction time domain, 5 min for control time domain, and 1 min for sampling period. These are all exemplary parameters and can be set according to actual engineering needs without changing the rolling optimization principle.

[0170] The experimental and comparative examples used the same wind and solar power disturbance sequence, ambient temperature boundary, load and grid conditions, total rated power and energy of energy storage, and initial SOC / SOE conditions. The comparative examples used a traditional single-battery energy storage scheme or a static power allocation strategy without spectrum decoupling and smooth transfer mechanisms; the examples used the flywheel-battery composite energy storage system and hierarchical coordinated control method proposed in this invention. The corresponding indicators were calculated based on synchronously sampled data of grid connection point power, grid frequency, DC bus voltage, and flywheel and battery branch power.

[0171] Figure 3 and Figure 4 In the diagram, A represents the grid-connected power fluctuation rate, B represents the maximum frequency deviation, C represents the proportion of high-rate impact energy from the battery, D represents the power command response time, E represents the disturbance power recovery time, and F represents the maximum DC bus voltage deviation.

[0172] like Figure 3 and Figure 4 As shown, traditional systems, lacking power-energy decoupling, exhibit grid-connected power fluctuations of 8.5%, a maximum frequency deviation of 82 MHz, a 31% proportion of high-rate battery impact energy, power command response and recovery times of 760 ms and 620 ms respectively, and a maximum DC bus voltage deviation of 46 V. This invention, through flywheel-battery composite energy storage and spectrum decomposition coordinated control, reduces grid-connected power fluctuations to 5%, compresses the maximum frequency deviation to 50 MHz, reduces the proportion of high-rate battery impact energy to 18%, shortens power command response and recovery times to 460 ms and 370 ms respectively, and suppresses the maximum DC bus voltage deviation to 27 V. These results demonstrate that this invention significantly outperforms traditional solutions in terms of power quality, battery life protection, dynamic response, and DC bus stability, effectively solving the problem of battery life degradation caused by high-power impacts in desert environments.

Claims

1. A flywheel-battery hybrid energy storage system, characterized in that, include: The energy management layer configures the power reference commands used to generate the flywheel and battery. The coordination and allocation layer is configured to perform power spectrum splitting on the power reference command, allocating high-frequency components to flywheel power commands and low-frequency components to battery power commands. It also identifies the flywheel's operating boundary in advance and smoothly transfers the excess portion to the battery at a defined slope before reaching the operating boundary; and The equipment control layer is configured to perform temperature adaptation control, including: temperature adaptation control of the flywheel to cope with high-frequency power fluctuations; In addition, temperature adaptation control of the battery is implemented to achieve energy time shift and power smoothing.

2. The flywheel-battery hybrid energy storage system according to claim 1, characterized in that, The energy management layer generates power reference commands by combining wind and solar ultra-short-term forecasts with grid status feedback, aiming to smooth the total power output of power plants and participate in grid frequency regulation, and then continuously optimizing to generate power reference commands for flywheels and batteries within a preset forecast step size in the future.

3. The flywheel-battery hybrid energy storage system according to claim 2, characterized in that, It also includes power electronic conversion and grid connection systems, which are used to aggregate the output power of the flywheel and battery to the DC bus, and finally connect to the grid or supply the load; The power electronic conversion and grid connection system generates grid status feedback based on the equipment status and the grid system status.

4. The flywheel-battery hybrid energy storage system according to claim 1, characterized in that, The high-frequency component is a power component with a change period of <1 minute, large amplitude fluctuation, and short duration; the low-frequency component is a power component with a change period of ≥1 minute, gradual change, and long duration.

5. The flywheel-battery hybrid energy storage system according to claim 1, characterized in that, The operating boundary is when the flywheel speed is close to 95% of the rated speed upper limit or 20% of the rated speed lower limit, or when the adjustable capacity is insufficient.

6. The flywheel-battery hybrid energy storage system according to claim 1, characterized in that, The device control layer performs temperature adaptation control of the flywheel in the following ways: In summer, under extreme high temperature conditions, the flywheel charging and discharging power limit is adjusted and the safe speed range is narrowed; in winter, under extreme low temperature conditions, the low-speed preheating logic is executed first, and the rated power capacity is released after the bearing temperature rises to a reasonable range.

7. The flywheel-battery hybrid energy storage system according to claim 6, characterized in that, The device control layer performs temperature adaptation control of the battery in the following ways: In extreme high-temperature environments in summer, the battery charge / discharge rate limit is lowered and the charge / discharge voltage window is narrowed; in extreme low-temperature environments in winter, the thermal management system is activated to preheat the battery cells, while the charge / discharge rate is dynamically lowered and the charging cutoff voltage is adjusted.

8. A method for hierarchical coordinated control of a flywheel-battery hybrid energy storage system based on any one of claims 1 to 7, characterized in that, Includes the following steps: Construct a power prediction model; A standard quadratic programming form is established, which includes an objective function and constraints. The objective function adopts a weighted summation of three objectives and minimizes each performance index. At the same time, multi-dimensional constraints are established, including power balance constraints, time-varying power limit constraints, state boundary constraints, and power change rate constraints. Among them, the upper and lower limits of the power in the time-varying power limit constraints are dynamically adjusted according to real-time temperature and state. The optimal control sequence is obtained by using a sequential quadratic programming algorithm. The first element of the optimal control sequence is used as the control variable at the current moment, which includes the battery power command and the flywheel power command. The battery power command and flywheel power command are sent to the coordination and allocation layer; at the same time, the measured changes in the state of charge and state of energy of the energy storage system are collected, and the model prediction error is calculated; if the error exceeds the set threshold, the prediction model parameters or initial state estimate are adjusted through state feedback correction in the next control cycle. Rolling time-domain update and iterative optimization.

9. The hierarchical coordination control method according to claim 8, characterized in that, The steps for constructing the power prediction model include: Multi-source feature extraction and input vector construction; A two-layer LSTM is used for temporal feature extraction; and a Dropout regularization mechanism is introduced after the output layer of the second LSTM; the feature vector after Dropout is linearly transformed through a fully connected layer and mapped to the prediction output space. The power prediction model is trained and optimized based on a loss function that incorporates a gust mutation penalty term; Backpropagation and parameter optimization; Output of ultra-short-term prediction sequences.

10. The hierarchical coordination control method according to claim 8, characterized in that, The performance indicators include a tracking error term, used to minimize the deviation between grid-connected power and the dispatch plan; a frequency support term, used to minimize the sum of squares of grid frequency deviations, dynamically increasing the frequency regulation weight when the frequency deviation exceeds the dead zone; and a lifetime degradation term.

Images