A power prediction method for power turbine based on disturbance observation and dynamic fusion

Abstract

The application discloses a power turbine power prediction method based on disturbance observation and dynamic fusion, carries out time alignment and normalization processing on input flow and actual output power of the power turbine, establishes a nominal physical model containing lumped disturbance, and extracts the lumped disturbance formed by unmodeled dynamics and external disturbance by using an extended state observer; the input flow, the actual output power and the input flow change amount are constructed into time sequence characteristics, a GRU agent model is trained to predict the lumped disturbance; a pure physical prediction channel and a disturbance compensation prediction channel are established during online prediction, the fusion weight is dynamically distributed in an inverse variance manner according to the local variance of historical prediction errors of the two channels, and the weight is smoothed through an exponential moving average, so that the fusion prediction power is obtained; the power prediction precision, dynamic adaptability and prediction stability of the power turbine under transient conditions can be improved.

Description

Technical Field

[0001] This invention relates to the field of engine waste heat recovery and power turbine power prediction technology, specifically to a power turbine power prediction method based on disturbance observation and dynamic fusion. Background Technology

[0002] Engine waste heat recovery systems typically utilize the residual energy in exhaust gas to drive a power turbine, thereby improving system energy efficiency. The power turbine output power is a crucial parameter reflecting the state of waste heat recovery and for energy management; its accurate prediction is essential for system operation analysis and control.

[0003] Existing methods for predicting power turbine output mainly include physics-based methods and data-driven methods. Physics-based methods typically establish prediction models based on thermodynamic relationships, state-space equations, or steady-state characteristics, which have clear physical meanings. However, in actual operation, power turbines are affected by factors such as rapid changes in input flow, model parameter deviations, unmodeled dynamics, and external disturbances. Fixed-parameter physical models are difficult to fully characterize dynamic changes under transient conditions, which can easily lead to prediction errors.

[0004] Data-driven methods can learn the nonlinear mapping relationship between input and output using historical operating data, but their prediction results depend on the distribution of training data and are difficult to maintain consistency with the physical dynamics of the power turbine. Existing physical-data fusion methods typically use the data model to directly learn the residual between the actual output and the prediction result of the physical model, and compensate according to a fixed structure or fixed ratio. When the operating conditions of the power turbine change, the reliability of the predictions of the physical model and the data model will also change, and fixed compensation or fixed fusion relationship is difficult to adjust in a timely manner.

[0005] In view of this, the present invention provides a method for predicting power of a power turbine based on disturbance observation and dynamic fusion. Summary of the Invention

[0006] The purpose of this invention is to provide a power turbine power prediction method based on disturbance observation and dynamic fusion, which solves the problems that existing power turbine power prediction methods are difficult to adapt to both unmodeled dynamic and transient disturbances at the same time, and that the prediction reliability is difficult to dynamically adjust with operating conditions when the physical model and the data-driven model adopt a fixed fusion relationship.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a method for predicting the power of a power turbine based on disturbance observation and dynamic fusion, comprising the following steps:

[0009] The input flow and actual output power during the historical operation of the power turbine are obtained, and the input flow and actual output power are time-aligned and normalized.

[0010] A nominal physical model of the power turbine is established based on the processed input flow and actual output power. Unmodeled dynamics and external disturbances that the nominal physical model fails to describe are identified as lumped disturbances. An extended state observer is constructed to estimate the lumped disturbances based on the input flow and actual output power.

[0011] Construct a time-series input sequence for continuous sampling times based on the input flow, actual output power, and changes in input flow. Use the lumped disturbance estimated by the extended state observer as the training target to train the GRU surrogate model, so that the GRU surrogate model can predict the lumped disturbance at the next sampling time based on the time-series input sequence obtained up to the current sampling time.

[0012] Online acquisition of power turbine operating data; based on the actual output power and input flow obtained up to the current sampling time, the pure physical power prediction value for the next sampling time is formed through the nominal physical model; and the lumped disturbance predicted by the GRU proxy model for the next sampling time is introduced into the nominal physical model as a compensation representation of unmodeled dynamics and external disturbances within the prediction interval, forming the data-compensated power prediction value for the next sampling time.

[0013] The local error variances of the two prediction channels are determined based on the historical prediction errors obtained up to the current sampling time. The fusion weights of the two prediction channels are dynamically determined according to the inverse variance relationship of the local error variances, so that the prediction channel with a relatively smaller local error variance receives a relatively higher fusion weight. The fusion weights are then smoothed by an exponential moving average.

[0014] The pure physical power prediction value and the data-compensated power prediction value are fused according to the smoothed fusion weights to obtain the power turbine fusion prediction power at the next sampling time. After obtaining the actual output power at the next sampling time, the historical prediction errors of the two prediction channels and the time-series input sequence are updated for lumped disturbance prediction and fusion weight update at subsequent sampling times.

[0015] Furthermore, the input flow rate and actual output power are time-aligned and normalized, including: resampling the input flow rate and actual output power using linear interpolation to map them to equally spaced time grids with the same sampling period; normalizing the resampled input flow rate and actual output power using Min-Max normalization and saving the corresponding normalization parameters so that real-time running data can be processed according to the same normalization parameters during the online prediction stage.

[0016] Furthermore, the nominal physical model adopts a first-order state-space model that includes the input flow rate and output power state; the lumped disturbance is expanded into a system state, and it is assumed that the lumped disturbance is differentiable; the expanded state observer is a linear expanded state observer, which updates the output power state estimate and the lumped disturbance estimate based on the observation error between the actual output power of the power turbine and the output power state estimate.

[0017] Furthermore, the observer gain of the linearly extended state observer is determined using a pole configuration method based on the preset observer bandwidth, wherein the first observer gain is twice the observer bandwidth and the second observer gain is the square of the observer bandwidth; the observer bandwidth is determined based on the frequency band characteristics of the power turbine system.

[0018] Furthermore, the change in input flow rate is determined based on the difference in input flow rate between adjacent sampling times, and the input flow rate, actual output power, and change in input flow rate at the same sampling time are used to form an extended feature vector. The extended feature vectors of multiple consecutive sampling times are extracted in chronological order to form the time-series input sequence, the end of the time-series input sequence corresponds to the current sampling time, and the lumped disturbance estimate of the next sampling time after the time-series input sequence is used as the corresponding training label.

[0019] Furthermore, the GRU proxy model includes a first GRU layer, a second GRU layer, and a fully connected output layer connected in sequence. The hidden state formed by the first GRU layer at each time step is used as the input of the second GRU layer at the corresponding time step. The hidden state formed by the second GRU layer at the end of the time-series input sequence is mapped to the lumped perturbation prediction value at the next sampling time through the fully connected output layer, and the mean square error between the lumped perturbation prediction value and the corresponding lumped perturbation estimate value is used as the training loss.

[0020] Furthermore, during the online prediction process, when the input flow is lower than the preset low flow threshold and continues to reach the preset number of sampling steps, lumped disturbance soft attenuation is triggered. Upon the first trigger, the lumped disturbance prediction value formed by the GRU proxy model is used as the initial value for soft attenuation. Subsequently, when the low flow condition is met continuously, the disturbance compensation value used by the data compensation prediction channel is recursively attenuated based on the disturbance compensation value after soft attenuation at the previous sampling time, according to an attenuation coefficient less than 1 and greater than 0. When the input flow reaches or exceeds the low flow threshold again, the soft attenuation is released and the lumped disturbance prediction value formed by the GRU proxy model is restored.

[0021] Furthermore, the pure physical power prediction value is determined according to the discrete forward relation of the nominal physical model based on the actual output power obtained at the current sampling time, the input flow rate at the current sampling time, the sampling period, and the nominal model parameters, and corresponds to the power turbine output power at the next sampling time; the data-compensated power prediction value is based on the same discrete forward relation with the addition of the lumped disturbance compensation value formed for the next sampling time. The lumped disturbance compensation value is the lumped disturbance predicted by the GRU surrogate model at the next sampling time under normal operating conditions, and is the disturbance compensation value after soft attenuation when low flow soft attenuation is triggered.

[0022] Furthermore, after obtaining the actual output power at each predicted sampling time, the historical prediction errors of the pure physical power prediction value and the data compensation power prediction value relative to the corresponding actual output power are determined, and a historical prediction error sliding window of a preset length is established for each. The local error variance is determined based on the degree of dispersion of the historical prediction error relative to the corresponding error mean within each sliding window, and the local error variance is used to characterize the fluctuation of the prediction error of the corresponding prediction channel in the most recent operating interval. When the number of effective historical prediction errors in the online prediction startup phase does not reach the preset window length, the local error variance is calculated based on the already obtained effective historical prediction errors, and a preset initial fusion weight is used when the effective historical prediction errors are insufficient to calculate the local error variance.

[0023] Furthermore, the initial fusion weights are determined based on the reciprocals of the local error variances of the two prediction channels, so that the prediction channel with relatively low recent prediction error fluctuations receives a higher initial fusion weight. The same smoothing factor is used to perform exponential moving average processing on the two initial fusion weights, and the sum of the two smoothed fusion weights is kept at 1. The power prediction values ​​formed by the two prediction channels for the same next sampling time are weighted and fused according to the smoothed fusion weights. When the power prediction value is formed in the normalized space, the fusion prediction result is de-normalized according to the actual output power normalization parameters saved in the offline stage to obtain the power turbine fusion prediction power with actual power dimensions.

[0024] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0025] This invention unifies the unmodeled dynamics and external disturbances that the nominal physical model of a power turbine fails to describe as lumped disturbances, and uses an extended state observer to estimate these lumped disturbances, thus separating the nominal dynamics that the physical model can describe from the dynamic deviations that need to be compensated. By using the estimated lumped disturbance as the training target, a GRU surrogate model is trained using a time-series input sequence containing input flow, actual output power, and changes in input flow. This enables the prediction of subsequent lumped disturbances based on transient changes in operating conditions and input conditions, improving adaptability to rapid changes in operating conditions. Furthermore, by establishing a pure physical prediction channel and a data-compensated prediction channel respectively, and dynamically allocating fusion weights according to the local variance of the recent historical prediction errors of the two prediction channels in an inverse variance manner, the channel with higher prediction performance receives a higher weight. At the same time, by smoothing the fusion weights using an exponential moving average, the sudden changes in weights caused by transient errors can be reduced, thereby improving the accuracy, stability, and dynamic adaptability of power turbine power prediction under different operating conditions. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0027] Figure 1 This is a schematic diagram of the power turbine power prediction method of the present invention. Detailed Implementation

[0028] The technical solution of the present invention will be further described below with reference to the accompanying drawings. The following embodiments are used to illustrate the technical principles and operation process of the present invention, and are not intended to limit the scope of protection of the present invention. Equivalent substitutions made by those skilled in the art to the execution subject, data storage format, model type, scheduling platform, or parameter configuration method without departing from the technical concept of the present invention shall all fall within the scope of protection of the present invention.

[0029] Example 1

[0030] like Figure 1 As shown, this embodiment provides a power prediction method for power turbines based on disturbance observation and dynamic fusion, applied to an engine-power turbine waste heat recovery and utilization system. The method includes an offline modeling stage and an online prediction stage.

[0031] In the offline modeling stage, a nominal physical model is established using historical operating data of the power turbine. The lumped disturbances that the nominal physical model fails to describe are extracted by an extended state observer, and a GRU surrogate model is established using the lumped disturbances as the training target.

[0032] In the online prediction phase, the trained GRU surrogate model is used to predict lumped disturbances. Pure physical prediction channels and data-compensated prediction channels are formed by the nominal physical model and the physical model after adding the predicted lumped disturbances, respectively. The local error variance is calculated based on the historical prediction errors of the two prediction channels. The fusion weights of the two prediction channels are dynamically determined based on the local error variances. The fusion weights are smoothed by exponential moving average. The power prediction results of the two prediction channels are fused based on the smoothed fusion weights to obtain the predicted power of the power turbine.

[0033] To unify the time relationship between the offline training phase and the online prediction phase, this embodiment uses the k-th sampling time as the time when the actual operating data has been obtained, and the (k+1)-th sampling time as the time to be predicted. During the online prediction process, only the input flow, actual output power, and historical prediction errors obtained up to the k-th sampling time are used to predict the lumped disturbance and power turbine output power at the (k+1)-th sampling time, without invoking the actual output power that has not yet been obtained at the (k+1)-th sampling time. After the actual output power at the (k+1)-th sampling time is obtained, the historical prediction errors of the two prediction channels are updated using this actual output power, and then used for dynamic fusion prediction at subsequent sampling times.

[0034] I. Offline Modeling Stage

[0035] First, acquire multi-source sensor data generated during the historical operation of the power turbine, including at least the original control input sequence. and output sequence .in, Input flow includes the exhaust mass flow rate entering the power turbine; This represents the actual output power of the power turbine, i.e., the sensor observation value at the corresponding moment.

[0036] Since the input flow and actual output power may originate from different data acquisition channels, their original sampling times may differ. Therefore, linear interpolation is used to resample the original control input and output sequences, mapping them uniformly to an equally spaced time grid. In this embodiment, the resampling period for the time interval between two adjacent sampling points is: .

[0037] After resampling, a Min-Max global normalization strategy is used to normalize the resampled input traffic. and actual output power They are mapped to the dimensionless interval [0,1] respectively to unify the data scale of different variables.

[0038] In the offline modeling phase, the Min-Max normalized parameters corresponding to the input flow and actual output power are saved separately. In the online prediction phase, the same normalized parameters are used to process the real-time input flow and actual output power to ensure that the offline training data and the online input data have a consistent data scale. The nominal physical model, LESO, GRU surrogate model, and subsequent dual-channel dynamic fusion process are all executed under the corresponding unified data scale.

[0039] After data preprocessing, a nominal physical model of the power turbine is established. Specifically, the first-order state-space equations of the power turbine are constructed, and the system dynamics are divided into nominal physical dynamics that can be described by the nominal physical model and unmodeled dynamics that cannot be directly described by the nominal physical model. The sum of the unmodeled dynamics and external disturbances is defined as the lumped disturbance, and the lumped disturbance is expanded into a new system state. Assuming it is differentiable, the state-space model of the power turbine expansion is expressed as:

[0040] ;

[0041] In the formula: This refers to the system state corresponding to the output power of the power turbine. For lumped disturbances; Let be the rate of change of the lumped disturbance, i.e., assuming the lumped disturbance is differentiable; This refers to the actual output power of the power turbine. For input traffic; and These are the system's nominal parameters, obtained through power turbine mechanism analysis or offline identification of historical operating data.

[0042] Based on the above extended state-space model, a linear extended state observer (LESO) is established to estimate the power output state and lumped disturbance of the power turbine, as shown in the following formula:

[0043] ;

[0044] In the formula, System status The estimated value; For aggregated disturbance The estimated value; This is the observation error; and This is the observer gain.

[0045] The observer gain is tuned based on pole placement and bandwidth parameterization:

[0046] ; ;

[0047] In the formula, The observer bandwidth is adjusted based on the system frequency band characteristics of the power turbine. This is achieved by continuously inputting pre-processed input flow. and actual output power LESO updates in real time and ,in, Used to estimate the output power state of the power turbine. Used to estimate lumped disturbances.

[0048] After obtaining the lumped disturbance estimation results, a GRU proxy model is constructed.

[0049] First, the input feature space is constructed. To characterize the transient changes in input flow, the change in input flow is added as a transient stimulus feature to the original input flow. For the k-th sampling time, the change in input flow is expressed as:

[0050] ;

[0051] In the formula, This represents the change in input flow rate at the current sampling time relative to the previous sampling time. and The first The sampling time and the first sampling time Input flow rate at each sampling time.

[0052] Construct an extended feature vector based on the input flow rate, actual output power, and change in input flow rate: ;

[0053] In the formula, For the first The extended feature vector at each sampling time.

[0054] Set the length of the sliding window Extracting consecutive sequences in chronological order The extended feature vectors at each sampling time point form the GRU input sequence: ;

[0055] The GRU input sequence The last extended feature vector in the sequence corresponds to the first... Therefore, the input sequence only contains data up to the sampling time. The actual operational data obtained at each sampling time.

[0056] The lumped perturbation estimate obtained from LESO at the next sampling time is used as the corresponding output label: ;

[0057] Therefore, the training relationship of the GRU surrogate model is determined as follows: up to the [number]th [year]... The continuous operating state at the nth sampling time is used as input, with the nth sampling time as input. The lumped disturbance estimate obtained by LESO at each sampling time is used as a supervision label, enabling the GRU surrogate model to learn the temporal mapping relationship between "historical and current operating state - lumped disturbance at the next sampling time".

[0058] This results in a time-series training sample that takes the continuous operating state as input and the corresponding lumped disturbance estimate as the training objective.

[0059] This embodiment employs a two-layer GRU network. The core idea of ​​GRU is to control the flow of information through a "gate" mechanism. The calculation process for each layer of GRU at time step t is as follows:

[0060] Reset door ;

[0061] Update Gate: ;

[0062] Candidate hidden state: ;

[0063] Final hidden state: ;

[0064] In the formula: For time steps Extended feature input vector; This is the hidden state from the previous time step; Use the Sigmoid activation function; It is the hyperbolic tangent activation function; This represents the Hadamard product (element-wise multiplication). and These are the weight matrix and bias vector for the corresponding gate, respectively.

[0065] Two-layer GRU architecture and final fully connected output: In the two-layer GRU architecture, the hidden state output of the first-layer GRU at each time step... As the input vector of the second-layer GRU at the corresponding time step .

[0066] make This indicates that the second-level GRU is at the end of the sliding window sequence (i.e., the first...). Each time step corresponds to the current moment. The final hidden state is formed. This hidden state... Input to a fully connected output layer to obtain the next sampling time (the... Lumped disturbance prediction at time (time) :

[0067] ;

[0068] In the formula: This represents the hidden state of the second-layer GRU at the end of the sliding window sequence; and These are the weight matrix and bias terms of the fully connected output layer, respectively; This is the lumped perturbation value predicted by the GRU surrogate model at the next sampling time.

[0069] Mean squared error is used as the training loss function: ;

[0070] In the formula, For training loss; The parameters to be trained for the GRU surrogate model; This represents the number of training samples; The lumped disturbance prediction value output by the GRU surrogate model; These are the corresponding lumped disturbance observations. Through training, the GRU surrogate model establishes a time-series mapping relationship between input flow, power turbine output state, input flow changes, and lumped disturbances.

[0071] For low-flow idling conditions, when the input flow rate satisfies: u(k) th And the state lasts for more than N. delay During each sampling step, the perturbation prediction value output by the GRU is softly attenuated: , ;

[0072] In the formula, u th The threshold for determining low flow rate; N delay The number of steps to continuously determine the low flow state; This is the attenuation coefficient. By continuously attenuating the lumped disturbance prediction value output by the GRU, the data compensation amount when the power turbine is continuously in a low flow state is gradually reduced. The soft attenuation is only enabled under the corresponding operating conditions of offline inference or online prediction and does not participate in the training of the GRU surrogate model.

[0073] In this embodiment, the soft attenuation effect is applied to the actual perturbation compensation value used in the data compensation prediction channel, rather than modifying the internal parameters of the GRU proxy model or retraining the GRU proxy model.

[0074] The purpose of using the above-mentioned soft attenuation is to gradually reduce the disturbance compensation amount in the data compensation prediction channel when the power turbine is continuously in a low flow state, so that the data compensation prediction result gradually converges to the prediction result determined by the nominal physical model, thereby avoiding maintaining a large historical disturbance compensation amount under continuous low flow conditions.

[0075] ​When the input flow reaches or exceeds the low flow threshold again, the soft attenuation process is released, and the original lumped disturbance prediction value formed by the GRU proxy model based on the current time-series input sequence is used as the disturbance compensation value of the data compensation prediction channel.

[0076] The soft attenuation is only enabled under the corresponding operating conditions in online prediction and is not involved in the training of the GRU surrogate model.

[0077] II. Online Prediction Phase

[0078] After completing offline modeling, online power prediction is performed using the nominal physical model and the trained GRU surrogate model. During the online prediction process, assume that the current value is the [number]th [unit]. Input flow rate at each sampling time and actual output power The prediction target is unified as the first The power output of the power turbine at each sampling moment.

[0079] For the current sampling time, the GRU input sliding window is updated based on the obtained operational data, and the GRU surrogate model outputs the lumped disturbance prediction value for the corresponding prediction time. A pure physical prediction channel and a data compensation prediction channel are then established based on the lumped disturbance prediction value.

[0080] The pure physics prediction channel uses only the nominal physics model for power prediction:

[0081] ;

[0082] In the formula, For pure physics prediction channels in the first... Power prediction values ​​obtained at each sampling time; This represents the actual output power of the power turbine at the previous sampling time. This represents the current input flow. and These are the nominal model parameters used for online prediction.

[0083] The data compensation prediction channel, based on the same nominal physical model, incorporates the lumped perturbation prediction value output by the GRU surrogate model as a compensation term into the physical equation:

[0084] ;

[0085] In the formula, The power prediction value obtained from the data compensation prediction channel; This is the lumped disturbance prediction value output by the GRU proxy model.

[0086] For the two prediction channels, a historical prediction error sliding window with a maintenance length of Nvar is used. Once the actual output power of the power turbine at the corresponding sampling time has been obtained, the prediction errors for the pure physics prediction channel and the data-compensated prediction channel are calculated separately:

[0087] ;

[0088] ;

[0089] In the formula, For pure physics prediction channels in the first... The prediction error at each sampling time; For data compensation prediction channels in the first The prediction error at each sampling time.

[0090] Based on the historical prediction errors of the most recent Nvar sampling times, calculate the local error variances of the two prediction channels respectively:

[0091] ;

[0092] ;

[0093] In the formula, This represents the local error variance of the purely physical prediction channel. To compensate for the local error variance of the prediction channel; and These represent the mean prediction error within the corresponding sliding window. The local error variance is used to represent the degree of fluctuation in the prediction error of the corresponding prediction channel within its most recent operating range.

[0094] The local error variance is used to characterize the fluctuation of the prediction error of the corresponding prediction channel within the most recent operating range. The smaller the local error variance, the lower the fluctuation of the recent prediction error of the corresponding prediction channel and the higher the prediction stability. Therefore, this embodiment does not directly equate the local error variance with the absolute magnitude of the prediction error or the prediction accuracy.

[0095] During the online prediction initiation phase, if the number of valid historical prediction errors obtained is less than Nvar, the corresponding local error variance is calculated using the currently obtained valid historical prediction errors. If the number of valid historical prediction errors is insufficient to calculate the local error variance, the two prediction channels use preset initial fusion weights. Once the number of valid historical prediction errors meets the conditions for calculating the local error variance, the system switches to a dynamic fusion method based on the local error variance.

[0096] The initial fusion weights are determined using the inverse variance method based on the local error variance of the two prediction channels:

[0097] ;

[0098] ;

[0099] In the formula, A numerical regularization smoothing factor is used to avoid overflow of the denominator due to excessively small error variance in extreme cases. The initial fusion weights for the pure physics prediction channels; The initial fusion weights are used to compensate for the data prediction channel.

[0100] The above expression is equivalent to the inverse variance weighted form, which calculates and normalizes the inverse variances of the local errors of the two channels respectively, and satisfies: ;

[0101] When the local error variance of a prediction channel increases relative to another prediction channel, the initial fusion weight obtained by that prediction channel decreases accordingly; when its local error variance decreases relatively, the corresponding initial fusion weight increases accordingly.

[0102] Therefore, the inverse variance weighting is used to dynamically adjust the channel fusion weights based on the fluctuation of the recent prediction errors of the two prediction channels, so that the channels with relatively lower fluctuations in recent prediction errors receive higher fusion weights.

[0103] To reduce abrupt changes in fusion weights caused by transient noise, the initial fusion weights of the two prediction channels are further smoothed using an exponential moving average:

[0104] ;

[0105] In the formula, and These are the fusion weights of the pure physical prediction channel and the fusion weights of the data compensation prediction channel, respectively, after exponential moving average processing. and These are the fusion weights corresponding to the previous sampling time; As a smoothing factor, The value can be 0.1 to 0.3.

[0106] When performing exponential moving average smoothing for the first time, the initial fusion weights obtained based on the current local error variance of the two prediction channels are used as the initial state of the EMA. When using preset initial fusion weights during the online prediction initiation phase, the sum of the two preset initial fusion weights is set to 1. Since both channels use the same smoothing factor α, and the sum of the two initial fusion weights input to the EMA is 1 each time, the two fusion weights remain unchanged after smoothing.

[0107] ;

[0108] Finally, based on the fusion weights smoothed by exponential moving average, the power prediction results generated by the pure physics prediction channel and the data compensation prediction channel for the next sampling time are weighted and fused:

[0109] ;

[0110] In the formula, For the first The fused predicted power of the dynamic turbine at each sampling time; For the pure physics prediction channel targeting the first Power prediction values ​​generated at each sampling time; For data compensation prediction channels targeting the first The power prediction value is generated at each sampling time.

[0111] When the After obtaining the actual output power of the power turbine at each sampling time, this actual output power is compared with the power prediction results generated by the two prediction channels for that sampling time. and The prediction errors are compared to obtain the updated channel prediction errors, and the new prediction errors are written into the corresponding historical prediction error sliding window. This updates the local error variance of the two prediction channels and the subsequent fusion weights.

[0112] Simultaneously, the newly acquired operational data is written into the GRU input sliding window, and the GRU surrogate model continues to predict the lumped disturbance at subsequent sampling times based on the updated input sequence. Thus, during the online prediction process, lumped disturbance prediction, pure physical power prediction, data-compensated power prediction, historical error update, local error variance calculation, fusion weight determination, EMA weight smoothing, and power fusion output are continuously executed.

[0113] The above process forms a continuous data transfer relationship: First, lumped disturbances are extracted from the actual operating data of the power turbine via ESO, and these lumped disturbances are used as the learning object of the GRU surrogate model; after the GRU surrogate model predicts the lumped disturbances based on the operating state of the power turbine, the predicted lumped disturbances are compensated to the nominal physical model to form a data compensation prediction channel, while retaining a pure physical prediction channel without disturbance compensation; the two prediction channels respectively form historical prediction errors based on the obtained actual output power, and the fusion weights are dynamically adjusted according to the local error variance within the sliding window, ultimately obtaining the fused predicted power of the power turbine. This data relationship is... Figure 1 The ESO perturbation extraction, GRU perturbation proxy modeling, and closed-loop dual-channel dynamic fusion structure shown correspond to these.

[0114] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A method for predicting the power of a power turbine based on perturbation observation and dynamic fusion, characterized in that, Includes the following steps: The input flow and actual output power during the historical operation of the power turbine are obtained, and the input flow and actual output power are time-aligned and normalized. A nominal physical model of the power turbine is established based on the processed input flow and actual output power. Unmodeled dynamics and external disturbances that the nominal physical model fails to describe are identified as lumped disturbances. An extended state observer is constructed to estimate the lumped disturbances based on the input flow and actual output power. Construct a time-series input sequence for continuous sampling times based on the input flow, actual output power, and changes in input flow. Use the lumped disturbance estimated by the extended state observer as the training target to train the GRU surrogate model, so that the GRU surrogate model can predict the lumped disturbance at the next sampling time based on the time-series input sequence obtained up to the current sampling time. Online acquisition of power turbine operating data; based on the actual output power and input flow obtained up to the current sampling time, the pure physical power prediction value for the next sampling time is formed through the nominal physical model; and the lumped disturbance predicted by the GRU proxy model for the next sampling time is introduced into the nominal physical model as a compensation representation of unmodeled dynamics and external disturbances within the prediction interval, forming the data-compensated power prediction value for the next sampling time. The local error variances of the two prediction channels are determined based on the historical prediction errors obtained up to the current sampling time. The fusion weights of the two prediction channels are dynamically determined according to the inverse variance relationship of the local error variances, so that the prediction channel with a relatively smaller local error variance receives a relatively higher fusion weight. The fusion weights are then smoothed by an exponential moving average. The pure physical power prediction value and the data-compensated power prediction value are fused according to the smoothed fusion weights to obtain the power turbine fusion prediction power at the next sampling time. After obtaining the actual output power at the next sampling time, the historical prediction errors of the two prediction channels and the time-series input sequence are updated for lumped disturbance prediction and fusion weight update at subsequent sampling times.

2. The method for predicting power turbine power based on disturbance observation and dynamic fusion according to claim 1, characterized in that, The input flow rate and actual output power are time-aligned and normalized, including: resampling the input flow rate and actual output power using linear interpolation and mapping them to equally spaced time grids with the same sampling period; normalizing the resampled input flow rate and actual output power using Min-Max normalization and saving the corresponding normalization parameters so that real-time running data can be processed according to the same normalization parameters during the online prediction stage.

3. The method for predicting power turbine power based on disturbance observation and dynamic fusion according to claim 1, characterized in that, The nominal physical model adopts a first-order state-space model that includes the input flow rate and output power state; the lumped disturbance is expanded into the system state, and it is assumed that the lumped disturbance is differentiable; the expanded state observer is a linear expanded state observer, which updates the output power state estimate and the lumped disturbance estimate based on the observation error between the actual output power of the power turbine and the output power state estimate.

4. The power prediction method for a power turbine based on disturbance observation and dynamic fusion according to claim 3, characterized in that, The observer gain of the linearly extended state observer is determined using a pole configuration method based on a preset observer bandwidth. The first observer gain is twice the observer bandwidth, and the second observer gain is the square of the observer bandwidth. The observer bandwidth is determined based on the frequency band characteristics of the power turbine system.

5. The method for predicting power turbine power based on disturbance observation and dynamic fusion according to claim 1, characterized in that, The change in input flow rate is determined based on the difference in input flow rate between adjacent sampling times. The input flow rate, actual output power, and change in input flow rate at the same sampling time are used to form an extended feature vector. The extended feature vectors of multiple consecutive sampling times are extracted in chronological order to form the time-series input sequence. The end of the time-series input sequence corresponds to the current sampling time, and the lumped disturbance estimate of the next sampling time after the time-series input sequence is used as the corresponding training label.

6. The power prediction method for a power turbine based on disturbance observation and dynamic fusion according to claim 5, characterized in that, The GRU proxy model includes a first GRU layer, a second GRU layer, and a fully connected output layer connected in sequence. The hidden state formed by the first GRU layer at each time step is used as the input of the second GRU layer at the corresponding time step. The hidden state formed by the second GRU layer at the end of the time-series input sequence is mapped to the lumped perturbation prediction value at the next sampling time through the fully connected output layer, and the mean square error between the lumped perturbation prediction value and the corresponding lumped perturbation estimate value is used as the training loss.

7. The method for predicting power turbine power based on disturbance observation and dynamic fusion according to claim 1, characterized in that, During online prediction, when the input flow is lower than the preset low flow threshold and continues to reach the preset number of sampling steps, lumped disturbance soft attenuation is triggered. Upon the first trigger, the lumped disturbance prediction value formed by the GRU proxy model is used as the initial value for soft attenuation. Subsequently, when the low flow condition is met continuously, the disturbance compensation value used by the data compensation prediction channel is recursively attenuated based on the disturbance compensation value after soft attenuation at the previous sampling time, according to an attenuation coefficient less than 1 and greater than 0. When the input flow reaches or exceeds the low flow threshold again, soft attenuation is released and the lumped disturbance prediction value formed by the GRU proxy model is restored.

8. The method for predicting power of a power turbine based on disturbance observation and dynamic fusion according to claim 1, characterized in that, The pure physical power prediction value is determined according to the discrete forward relation of the nominal physical model based on the actual output power obtained at the current sampling time, the input flow rate at the current sampling time, the sampling period, and the nominal model parameters, and corresponds to the power turbine output power at the next sampling time. The data-compensated power prediction value is based on the same discrete forward relation, with the addition of the lumped disturbance compensation value formed for the next sampling time. The lumped disturbance compensation value is the lumped disturbance predicted by the GRU surrogate model at the next sampling time under normal operating conditions, and is the disturbance compensation value after soft attenuation when low flow soft attenuation is triggered.

9. The method for predicting power turbine power based on disturbance observation and dynamic fusion according to claim 1, characterized in that, After obtaining the actual output power at each predicted sampling time, the historical prediction errors of the pure physical power prediction value and the data compensation power prediction value relative to the corresponding actual output power are determined, and a historical prediction error sliding window of a preset length is established for each. The local error variance is determined based on the degree of dispersion of the historical prediction error relative to the corresponding error mean within each sliding window. The local error variance is used to characterize the fluctuation of the prediction error of the corresponding prediction channel in the most recent operating interval. When the number of effective historical prediction errors in the online prediction startup phase does not reach the preset window length, the local error variance is calculated based on the already obtained effective historical prediction errors, and a preset initial fusion weight is used when the effective historical prediction errors are insufficient to calculate the local error variance.

10. The method for predicting power turbine power based on disturbance observation and dynamic fusion according to claim 9, characterized in that, Initial fusion weights are determined based on the reciprocals of the local error variances of the two prediction channels, so that prediction channels with relatively low recent prediction error fluctuations receive higher initial fusion weights. The same smoothing factor is used to perform exponential moving average processing on the two initial fusion weights, and the sum of the two smoothed fusion weights is kept to 1. The power prediction values ​​formed by the two prediction channels for the same next sampling time are weighted and fused according to the smoothed fusion weights. When the power prediction value is formed within the normalized space, the fusion prediction result is de-normalized according to the actual output power normalization parameters saved in the offline stage to obtain the power turbine fusion prediction power with actual power dimensions.

Images