Generator set operating parameter prediction method, device and storage medium based on multi-scale time series data fusion model

Abstract

The present invention discloses a method, device and storage medium for predicting operating parameters of a generator set based on a multi-scale time series data fusion model. The multi-scale time series data fusion model includes a data denoising module and a data fusion module. The data denoising module includes a denoising autoencoder, and the data fusion module includes a one-dimensional deep convolutional neural network and a bidirectional long short-term memory network. The one-dimensional deep convolutional neural network includes a convolutional layer, a pooling layer and a fully connected layer, and the bidirectional long short-term memory network is embedded in the one-dimensional deep convolutional neural network. The method for predicting the operating parameters of the generator set includes: collecting real-time data of each operating parameter affecting the operating mode of the generator set and performing preprocessing; inputting the preprocessed real-time data of each operating parameter into a pre-trained multi-scale time series data fusion model to obtain prediction results of each operating parameter of the generator set. The present invention can reduce the noise of the original data, improve the data fusion accuracy and the prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular, to a method, device and storage medium for predicting operating parameters of a generator set based on a multi-scale time-series data fusion model. Background Art

[0002] The purpose of data fusion is to obtain more accurate prediction data by fusing data captured by multiple sensors with information in a relevant database; currently, data fusion technology has been applied to multiple fields such as intelligent hydropower, intelligent transportation, disease diagnosis, stock prediction, etc., and the processed data belongs to time-series data, which often has characteristics such as high dimension, high noise, multiple features, continuity, and high non-linearity.

[0003] During the operation of cascade hydropower stations, reasonably arranging the operation mode and water volume scheduling of hydropower stations can enable hydropower stations and the power grids they are connected to to obtain the greatest possible economic benefits. However, being able to accurately predict operating parameters such as the flow rate, head, rotational speed, and power of each generator set plays a crucial role in the intelligent forecasting and reasonable scheduling of cascade hydropower stations, which is beneficial to reducing the total power consumed by each generator set and improving power generation efficiency. Currently, the main data prediction methods are to apply algorithms such as BP (Back Propagation) neural network, deep convolutional neural network, CNN-LSTM (Convolutional Neural Networks-Long Short-Term Memory) neural network, etc. to data fusion technology for prediction; however, the network structure of the BP neural network data fusion model is relatively simple, and it cannot achieve high-performance fusion efficiency for processing high-complexity data; the deep convolutional neural network data fusion model can well learn the characteristics of high-dimensional time-series data, but its model ignores the influence of data noise and cannot capture the continuity characteristics between time-series data; the CNN-LSTM neural network data fusion model solves the long-distance dependence problem between time-series data, but this model can only capture the continuity feature data of the forward input data, and the data fusion accuracy is not high; in summary, how to reduce the noise interference of the original data and achieve higher fusion accuracy without excessively increasing the model operation speed is an issue that researchers need to overcome. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies in the prior art, and provide a method, device and storage medium for predicting operating parameters of a generator set based on a multi-scale time-series data fusion model, which can reduce the noise of the original data, improve the data fusion accuracy and prediction accuracy, thereby accurately predicting the operating parameters of each generator set, and further being able to reasonably arrange and schedule the operation mode of cascade hydropower stations.

[0005] To achieve the above object, the present invention is implemented by the following technical solutions:

[0006] In a first aspect, the present invention provides a method for predicting the operating parameters of a generator set based on a multi-scale time series data fusion model. Among them, the multi-scale time series data fusion model includes a data denoising module and a data fusion module. The data denoising module includes a denoising autoencoder, and the data fusion module includes a one-dimensional deep convolutional neural network and a bidirectional long short-term memory network. The one-dimensional deep convolutional neural network includes a convolutional layer, a pooling layer, and a fully connected layer, and the bidirectional long short-term memory network is embedded in the one-dimensional deep convolutional neural network;

[0007] The method for predicting the operating parameters of the generator set includes:

[0008] Collect real-time data of each operating parameter affecting the operation mode of the generator set and perform preprocessing;

[0009] Input the real-time data of each preprocessed operating parameter into a pre-trained multi-scale time series data fusion model to obtain the prediction results of each operating parameter of the generator set.

[0010] Combined with the first aspect, preferably, each of the operating parameters includes the flow rate, head, rotational speed, and power of the generator set.

[0011] Combined with the first aspect, preferably, the training process of the multi-scale time series data fusion model includes the following steps:

[0012] Step 2.1: Initialize the model parameters of the multi-scale time series data fusion model;

[0013] Step 2.2: Collect historical data of each operating parameter of the generator set at different time scales, and the denoising autoencoder updates the collected historical data after denoising and reconstruction to obtain clean time series data;

[0014] Step 2.3: The one-dimensional deep convolutional neural network extracts features from the time series data to obtain a feature map containing spatial features and short-term time features;

[0015] Step 2.4: After performing a Flatten operation on the feature map, input it into the bidirectional long short-term memory network to further extract the long-term time features of the time series data in the feature map;

[0016] Step 2.5: The fully connected layer in the one-dimensional deep convolutional neural network fuses the spatial features, short-term time features, and long-term time features of the time series data, and the obtained fusion value is used as the prediction value of the model;

[0017] Step 2.6: Repeat Steps 2.2 to 2.5 for iterative training and update the network parameters of the model simultaneously until the value of the loss function of the model meets the set requirements, and then stop the training to obtain the trained multi-scale time series data fusion model.

[0018] Combined with the first aspect, preferably, the model parameters for initializing the multi-scale time series data fusion model include:

[0019] Collect the operation parameter data of the generator sets of different hydropower stations to pre-train the model;

[0020] Use the model parameters obtained after pre-training as the initial parameters of the multi-scale time series data fusion model.

[0021] Combined with the first aspect, preferably, the method for the denoising autoencoder to perform denoising reconstruction on the collected historical data includes the following steps:

[0022] Step a: Perform standard normalization processing on the collected historical data;

[0023] Step b: The input layer in the denoising autoencoder adds noise to the data after the normalization processing to obtain noisy input data;

[0024] Step c: Map the noisy input data to the hidden layer in the denoising autoencoder through the encoding function to obtain a feature representation;

[0025] Step d: Use the decoding function to map the obtained feature representation to the output layer in the denoising autoencoder;

[0026] Step e: Repeat Steps b to d, and iteratively update the parameters of the denoising autoencoder through the minimum cost function until the function converges, and reconstruct clean time series data.

[0027] Combined with the first aspect, preferably, the method for the one-dimensional depth convolutional neural network to extract features from the clean time series data includes:

[0028] The convolutional layer performs a convolution operation on the time series data to extract features, and the pooling layer compresses and reduces the dimension of the extracted features to reduce the computational amount of the network;

[0029] Among them, the calculation formula for the convolutional layer to extract features is:

[0030] Z l =σ(Z l-1 W l +b l ) (1)

[0031] In the formula, l represents the number of layers of the convolutional layer, Z l-1represents the feature map output by the (l - 1)-th layer, W l and b l are the weight and bias of the convolution kernel of the l-th layer respectively, σ represents the activation function, and Z l represents the feature map output by the l-th layer, and the feature map includes the spatial features and short-term temporal features of the input temporal data.

[0032] Combined with the first aspect, preferably, the long short-term memory network includes a forward long short-term memory network and a backward long short-term memory network.

[0033] Combined with the first aspect, preferably, the long short-term memory network extracts the long-term temporal features of the temporal data in the feature map through the calculation formula (2):

[0034]

[0035] In the formula, x t represents the input data of the long short-term memory network at time t; is the output result of the hidden layer state of the forward long short-term memory network at time t, is the output result of the hidden layer state of the forward long short-term memory network at time t - 1; is the output result of the hidden layer state of the backward long short-term memory network at time t, is the output result of the hidden layer state of the backward long short-term memory network at time t - 1; the output result of the hidden layer state of the forward long short-term memory network and the output result of the hidden layer state of the backward long short-term memory network jointly determine the long-term temporal features extracted by the long short-term memory network at time t.

[0036] In a second aspect, the present invention provides a generator set operation parameter prediction device based on a multi-scale temporal data fusion model, including a processor and a storage medium;

[0037] The storage medium is used to store instructions;

[0038] The processor is used to operate according to the instructions to execute the steps of the generator set operation parameter prediction method based on the multi-scale temporal data fusion model as described in any one of the first aspects.

[0039] In a third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the generator set operation parameter prediction method based on the multi-scale temporal data fusion model as described in any one of the first aspects are implemented.

[0040] Compared with the prior art, the beneficial effects achieved by the present invention:

[0041] The multi-scale time series data fusion model provided by the present invention first uses a denoising autoencoder to perform denoising reconstruction on input data of different time scales to solve the problems of high noise and partial data loss in the original data. Then, a one-dimensional deep convolutional neural network is used to extract the spatial features and short-term time features of the time series data after denoising reconstruction. At the same time, a bidirectional long short-term memory network is introduced to further obtain the long-term time features of the time series data, so as to improve the deficiency that only the continuity features of the forward input data can be captured in the traditional data fusion model. Finally, a fully connected layer is used to fuse the features of different scales extracted, enabling the model to output more accurate prediction results. This model can effectively solve the problems of high noise and poor fusion accuracy of the original data, enhancing the accuracy and robustness of data fusion. In addition, the method for predicting the operating parameters of a generator set using this model of the present invention can fully learn the potential change laws and mutual influence mechanisms between the data of the generator set, can effectively fuse the data under different time scales, and then obtain more accurate prediction results. Based on the accurate prediction of the data, the cascade hydropower station adopts reasonable operating and scheduling methods for the generator set to improve the overall operating efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 FIG. is a schematic flowchart of a method for predicting the operating parameters of a generator set based on a multi-scale time series data fusion model provided by an embodiment of the present invention;

[0043] Figure 2 FIG. is a schematic structural principle diagram of a multi-scale time series data fusion model provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] The technical solution of the present invention will be described in detail below with reference to the drawings and specific embodiments. It should be understood that the specific features in the embodiments of the present application and the embodiments are detailed descriptions of the technical solution of the present application, rather than limitations on the technical solution of the present application. Without conflict, the technical features in the embodiments of the present application and the embodiments can be combined with each other.

[0045] The term "and / or" in this article is only a description of the association relationship between associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " in this article generally represents an "or" relationship between the associated objects before and after.

[0046] Embodiment 1:

[0047] Referring to Figure 1 , an embodiment of the present invention introduces a method for predicting the operating parameters of a generator set based on a multi-scale time series data fusion model, which specifically includes the following steps:

[0048] Step 1: Collect the real-time data of each operating parameter affecting the operation mode of the generator set and perform preprocessing;

[0049] Step 2: Input the real-time data of each preprocessed operating parameter into a pre-trained multi-scale time series data fusion model to obtain the prediction results of each operating parameter of the generator set;

[0050] Among them, the predicted operating parameters include the flow rate, head, rotation speed, and power of the generator set. As an embodiment of the present invention, the structural composition of the multi-scale time series data fusion model provided by the embodiment of the present invention is as Figure 2 shown. This model includes a data denoising module and a data fusion module. Specifically, the data denoising module includes a denoising autoencoder, and the autoencoder includes an input layer, an output layer, and a hidden layer; the data fusion module includes a one-dimensional deep convolutional neural network and a bidirectional long short-term memory network. The one-dimensional deep convolutional neural network includes 4 convolutional layers, 2 pooling layers, and 1 fully connected layer; among them, the bidirectional long short-term memory network is embedded in the one-dimensional deep convolutional neural network.

[0051] The training process of the multi-scale time series data fusion model in Step 2 provided by the embodiment of the present invention includes the following steps:

[0052] Step 2.1: Initialize the model parameters of the multi-scale time series data fusion model;

[0053] It should be noted that in this step, the embodiment of the present invention pre-trains the model by collecting a large amount of generator set operation parameter data of different hydropower stations; among them, the collected data includes the power, flow rate, head, and rotation speed of each motor of the unit in a continuous time period; the purpose of pre-training is to enable the shallow network of the model to fully learn the potential change laws and mutual influence mechanisms between these data; the model parameters obtained after the pre-training are used as the initial parameters of each network in the model;

[0054] Step 2.2: Collect the historical data of each operating parameter of the generator set at different time scales, and the denoising autoencoder updates the collected historical data to obtain clean time series data after denoising and reconstruction;

[0055] Step 2.3: The one-dimensional deep convolutional neural network extracts features from the time series data to obtain a feature map containing spatial features and short-term time features;

[0056] Step 2.4: After performing a flatten operation on the feature map, input it into the bidirectional long short-term memory network to further extract the long-term time features of the time series data in the feature map;

[0057] Step 2.5: The fused value obtained by the fully connected layer in the one-dimensional depth convolutional neural network after fusing the spatial features, short-term time features, and long-term time features of the time series data is used as the predicted value of the model;

[0058] Step 2.6: Repeat Steps 2.2 to 2.5 for iterative training and simultaneously update the network parameters of the model until the loss function value of the model meets the set requirements, and then stop training to obtain a trained multi-scale time series data fusion model.

[0059] The method for denoising and reconstructing the collected historical data by the denoising autoencoder in Step 2.2 provided by the embodiments of the present invention includes the following steps:

[0060] Step a: Perform standard normalization processing on the collected historical data;

[0061] In this step, since different data have different measurement scales, for the convenience of model training, we perform standard normalization processing on the original input data to uniformly transform it into data in the range of [0, 1];

[0062] Step b: The input layer in the denoising autoencoder performs noise addition processing on the normalized data to obtain noisy input data;

[0063] Step c: Map the noisy input data to the hidden layer in the denoising autoencoder through an encoding function to obtain a feature representation;

[0064] Step d: Use a decoding function to map the obtained feature representation to the output layer in the denoising autoencoder;

[0065] Step e: Repeat Steps b to d, and iteratively update the parameters of the denoising autoencoder through a minimum cost function until the function converges. Take the network parameters obtained at this time as the network parameters of the autoencoder, and reconstruct clean time series data;

[0066] As an embodiment of the present invention, the minimization cost function minC of formula (1) is used as the loss function of the data denoising module:

[0067]

[0068] The cross-entropy L of formula (2) is used as the loss function for training the multi-scale time series data fusion model:

[0069]

[0070] In the formula, n represents the number of samples formed by the collected data, m is the size of the minimum batch, y i is the true value of the i-th sample, y^i is the output value of the i-th sample of the data denoising module, b i is the predicted value of the i-th sample; Adam is used as the optimizer for the data fusion module.

[0071] Furthermore, in step 2 of this embodiment, the historical data collected is original time series data of different scales. After standard normalization processing, it is uniformly mapped to the [0,1] interval, Gaussian white noise is added to it, and then it is input into the denoising autoencoder network for decoding and encoding. The method of unsupervised learning is used to fine-tune the network parameters so as to obtain the optimal parameter settings, thereby eliminating the noise of the original input data and obtaining clean output data, solving the defect of high noise in the original data of the traditional data fusion model.

[0072] The method for the one-dimensional deep convolutional neural network to extract features from the clean time series data in step 2.3 provided by the embodiment of the present invention includes:

[0073] The convolutional layer performs a convolutional operation on the time series data to extract features, and the pooling layer compresses and reduces the dimension of the extracted features to reduce the computational amount of the network;

[0074] Among them, the calculation formula for the convolutional layer to extract features is:

[0075] Z l =σ(Z l-1 W l +b l ) (3)

[0076] In the formula, l represents the number of layers of the convolutional layer, Z l-1 represents the feature map output by the (l-1)-th layer, W l and b l are respectively the weight and bias of the convolutional kernel of the l-th layer, σ represents the activation function, Z l represents the feature map output by the l-th layer, and the feature map includes the spatial features and short-term time features of the input time series data;

[0077] Further explanation is as follows. In step 3.2, the output data obtained in step 2.2 is used as the input of the data fusion module. A BN (Batch Normalization) layer is added after each convolutional layer to normalize the data, and the processed data is used as the input of the next layer. Each convolutional layer uses a convolutional kernel of size 1×5, and the activation function is ReLU. The input data of the data fusion module passes through two convolutional layers with padding of same and obtains an output of 7×35 dimensions. The output matrix is input into a pooling layer with a pooling window size of 1×3. The output of the max pooling layer is input into a one-dimensional convolutional layer with the same parameters as the previous two convolutional layers, and an output of 5×35 dimensions is obtained. After passing through four convolutional layers and two pooling layers with a pooling window size of 1×3 and a stride of 1, a feature map of size 3×35 is obtained. This feature map contains multi-dimensional features such as the space and short-term time of the input data.

[0078] As an embodiment of the present invention, the long short-term memory network in step 2.4 is composed of a forward long short-term memory network and a backward long short-term memory network combined; the long short-term memory network extracts the long-term time features of the time series data in the feature map through the calculation formula (4):

[0079]

[0080] In the formula, x t represents the input data of the long short-term memory network at time t; is the output result of the hidden layer state of the forward long short-term memory network at time t, is the output result of the hidden layer state of the forward long short-term memory network at time t-1; is the output result of the hidden layer state of the backward long short-term memory network at time t, is the output result of the hidden layer state of the backward long short-term memory network at time t-1; the output result of the hidden layer state of the forward long short-term memory network and the output result

[0081] of the hidden layer state of the backward long short-term memory network jointly determine the long-term time features extracted by the long short-term memory network at time t;

[0082]

[0083] In the formula, Γ f 、Γ u and Γ o represent the calculated values of the forgetting gate, update gate and output gate respectively; and c t are the candidate value and the updated value of the memory cell; σ and tanh represent the sigmoid function and the hyperbolic tangent function respectively; W c and b c represent the weight and bias of this unit respectively, W f 、W u 、W o and b f 、b u 、b o represent the weights and biases of the forget gate, update gate and output gate respectively; a t is the final output of this unit;

[0084] Furthermore, in step 2.4 provided in this embodiment, the output obtained in step 2.3 is tiled and transposed along the time dimension and then sent into a bidirectional long short-term memory network to capture the long-term time features of the time series data bidirectionally, solving the defect that the traditional data fusion model can only capture the continuity features between time series data unidirectionally.

[0085] As an embodiment of the present invention, in step 2.5 provided in the embodiment of the present invention, the output data obtained after being processed in step 2.4 is sent into a max pooling layer, and after being processed, it is input into a fully connected layer. The activation function of the fully connected layer is LeakyReLU. This activation function not only retains the advantages of the ReLU activation function, but also, even if the input value is negative, its derivative value will not be 0 and backpropagation can still be performed. The output of the max pooling layer is converted into the form of a one-dimensional vector, thereby effectively fusing the multi-dimensional features of the time series data, reducing the dimension of the data, and finally converting the output into the dimension of the original input data to obtain the fusion value, realizing the accurate fusion and prediction of the data of each index of the unit.

[0086] In summary, a method for predicting the operating parameters of a generator set based on a multi-scale time series data fusion model provided in the embodiment of the present invention fills the defects that the network structure of the traditional data fusion model is too simple, ignores the influence of the original data noise, and can only capture the continuity features between time series data unidirectionally; the multi-scale time series data fusion model constructed in the embodiment of the present invention proposes two improvements for the defects of high data noise and poor fusion accuracy of the existing similar models, and applies them to the intelligent forecasting and scheduling of cascade hydropower stations; compared with the traditional data fusion model, it has higher generalization ability and adaptability, higher data credibility, reduces the noise of the original data, improves the data fusion accuracy and prediction accuracy, can accurately predict the operating parameters of each generator set, and thus can reasonably arrange and schedule the operating mode of the cascade hydropower station, and is applicable to improving the overall efficiency of the hydropower station unit and the technical improvement of similar methods.

[0087] Embodiment 2:

[0088] An embodiment of the present invention provides a generator set operating parameter prediction device based on a multi-scale time series data fusion model, including a processor and a storage medium;

[0089] The storage medium is used to store instructions;

[0090] The processor is configured to operate according to the instructions to execute the steps of any one of the methods in the first embodiment.

[0091] Embodiment Three:

[0092] An embodiment of the present invention provides a computer-readable storage medium, on which a computer program is stored, characterized in that when the computer program is executed by a processor, the steps of any one of the methods in the first embodiment are implemented.

[0093] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0094] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0095] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and the instruction device implements the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0096] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so that a series of operation steps are executed on the computer or other programmable apparatus to produce a computer-implemented process, thereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the functions specified in one process or a plurality of processes and / or boxes Figure 1 one process or a plurality of processes and / or boxes Figure 1 steps for implementing the functions specified in one box or a plurality of boxes.

[0097] The above are only the preferred embodiments of the present invention. It should be pointed out that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A method for predicting operating parameters of a generator set based on a multi-scale time series data fusion model, characterized in that, the multi-scale time series data fusion model includes a data denoising module and a data fusion module. The data denoising module includes a denoising autoencoder, and the data fusion module includes a one-dimensional deep convolutional neural network and a bidirectional long short-term memory network. The one-dimensional deep convolutional neural network includes a convolutional layer, a pooling layer, and a fully connected layer, and the bidirectional long short-term memory network is embedded in the one-dimensional deep convolutional neural network; the method for predicting operating parameters of the generator set includes: collecting real-time data of each operating parameter affecting the operating mode of the generator set and performing preprocessing; inputting the preprocessed real-time data of each operating parameter into a pre-trained multi-scale time series data fusion model to obtain prediction results of each operating parameter of the generator set; the training process of the multi-scale time series data fusion model includes the following steps: Step 2.1: Initialize the model parameters of the multi-scale time series data fusion model; Step 2.2: Collect historical data of each operating parameter of the generator set at different time scales, and the denoising autoencoder updates the collected historical data after denoising and reconstruction to obtain clean time series data; Step 2.3: The one-dimensional deep convolutional neural network extracts features from the time series data to obtain a feature map containing spatial features and short-term time features; Step 2.4: After performing a Flatten operation on the feature map, input it into the bidirectional long short-term memory network to further extract the long-term time features of the time series data in the feature map; Step 2.5: The fully connected layer in the one-dimensional deep convolutional neural network fuses the spatial features, short-term time features, and long-term time features of the time series data, and the fused value obtained is used as the prediction value of the model; Step 2.6: Repeat Step 2.2 to Step 2.5 for iterative training and simultaneously update the network parameters of the model until the loss function value of the model meets the set requirements, then stop training to obtain a trained multi-scale time series data fusion model; the method for the denoising autoencoder to perform denoising and reconstruction on the collected historical data includes the following steps: Step a: Perform standard normalization processing on the collected historical data; Step b: The input layer in the denoising autoencoder adds noise to the normalized data to obtain noisy input data; Step c: Map the noisy input data to the hidden layer in the denoising autoencoder through an encoding function to obtain a feature representation; Step d: Use a decoding function to map the obtained feature representation to the output layer in the denoising autoencoder; Step e: Repeat Step b to Step d, and iteratively update the parameters of the denoising autoencoder through a minimum cost function until the function converges, and reconstruct to obtain clean time series data.

2. The method for predicting operating parameters of a generator set based on a multi-scale time series data fusion model according to claim 1, characterized in that, each of the operating parameters includes the flow rate, head, rotational speed, and power of the generator set.

3. The method for predicting the operating parameters of a generator set based on a multi-scale time series data fusion model according to claim 1, characterized in that, the model parameters for initializing the multi-scale time series data fusion model include: collecting the operating parameter data of the generator sets of different hydropower stations to pre-train the model; using the model parameters obtained after pre-training as the initial parameters of the multi-scale time series data fusion model.

4. The method for predicting the operating parameters of a generator set based on a multi-scale time series data fusion model according to claim 1, characterized in that, the method for the one-dimensional deep convolutional neural network to extract features from the clean time series data includes: the convolutional layer performs a convolution operation on the time series data to extract features, and the pooling layer compresses and reduces the dimension of the extracted features to reduce the computational amount of the network; wherein, the calculation formula for the convolutional layer to extract features is: （1） Wherein, represents the number of convolutional layers, represents the feature map of the output of the -th layer, and are the weight and bias of the convolutional kernel of the -th layer respectively, represents the activation function, represents the feature map of the output of the -th layer, and the feature map includes the spatial features and short-term temporal features of the input temporal data.

5. The method for predicting the operating parameters of a generator set based on a multi-scale time series data fusion model according to claim 1, characterized in that, the long short-term memory network includes a forward long short-term memory network and a backward long short-term memory network.

6. The method for predicting the operating parameters of a generator set based on a multi-scale time series data fusion model according to claim 5, characterized in that, the long short-term memory network extracts the long-term time features of the time series data in the feature map through the calculation formula (2): （2） In the formula, represents the input data of the long short-term memory network at a certain moment; is the output result of the hidden layer state of the forward long short-term memory network at a certain moment, is the output result of the hidden layer state of the forward long short-term memory network at a certain moment; is the output result of the hidden layer state of the backward long short-term memory network at a certain moment, is the output result of the hidden layer state of the backward long short-term memory network at a certain moment; the output result of the hidden layer state of the forward long short-term memory network and the output result of the hidden layer state of the backward long short-term memory network jointly determine the long-term time features extracted by the long short-term memory network at a certain moment.

7. A device for predicting the operating parameters of a generator set based on a multi-scale time series data fusion model, characterized in that, it includes a processor and a storage medium; the storage medium is used for storing instructions; the processor is used to operate according to the instructions to execute the steps of the method for predicting the operating parameters of a generator set based on a multi-scale time series data fusion model according to any one of claims 1 to 6.

8. A computer-readable storage medium, on which a computer program is stored, characterized in that, when the computer program is executed by a processor, it implements the steps of the method for predicting the operating parameters of a generator set based on a multi-scale time series data fusion model according to any one of claims 1 to 6.

Images