Deep learning industrial time series data prediction algorithm based on trend and residual error separation
Through a deep learning algorithm based on trend and residual separation, the time series data and control quantity data are decomposed into seasonal and trend components, feature extraction and dimensionality reduction are performed, and parallel calculations are performed to generate prediction values, which solves the problem of long recursive prediction time and accumulated errors, and improves prediction accuracy and adaptability.
Patent Information
- Application Number
- CN202510951139.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-10
- Publication Date
- 2025-08-08
AI Technical Summary
The existing technology relies on recursive methods for multi-step prediction, with a long training time and cumulative errors, resulting in poor accuracy.
Deep learning industrial timing data prediction algorithm based on trend and residual separation is adopted, and the timing data and control quantity data are decomposed into seasonal components and trend components through moving average and trend components, and feature extraction and dimensionality reduction are performed. Full connection layer and nonlinear activation functions are input to generate hidden representations, and finally training is carried out through backpropagation algorithm to minimize mean square error.
Parallel calculation is implemented, which avoids the accumulation of recursive errors, significantly shortens training time, improves prediction accuracy and ability to adapt to complex timing modes, and is suitable for large-scale timing prediction tasks.
Smart Images
Figure CN120448754A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of industrial big data computing, and in particular to a deep learning industrial time series data prediction algorithm based on trend and residual separation. Background Art
[0002] The application of time-series prediction algorithms holds significant strategic importance in the hydropower sector. Hydropower plant operations are influenced by numerous factors, including water flow, precipitation, temperature, and equipment status, often exhibiting significant time-series characteristics. Accurate time-series prediction can help hydropower plants better schedule generators, optimize reservoir storage and release strategies, improve power generation efficiency and safety, ensure sufficient power supply during peak hours, and avoid excessive water consumption. Furthermore, hydropower plant equipment often operates in complex environments, and the operational status of key equipment such as turbines and generators is crucial to power generation efficiency. By monitoring equipment status in real time and analyzing historical data, time-series prediction can predict equipment failure risks and schedule maintenance in advance, thereby avoiding downtime losses caused by unexpected failures and improving equipment reliability and availability. In terms of energy scheduling, hydropower plants often collaborate with other power supply sources. Especially in highly variable climate conditions, predicting trends in precipitation and water flow is crucial for the optimal scheduling of grid resources. By deeply mining time-series data from hydropower generation, more accurate load and power demand forecasts can be achieved, providing a scientific basis for grid scheduling and load balancing, and reducing the instability and risks associated with power supply fluctuations. With the development of intelligent technology, hydropower plants can optimize water resource management, equipment maintenance, and energy scheduling through time-series prediction algorithms, improving power generation efficiency and promoting sustainable development. In short, the application of time-series prediction in hydropower generation can significantly improve energy efficiency, ensure stable power plant operation, and contribute to the utilization of green energy and environmental protection.
[0003] Currently, most forecasting algorithms with control variables are based on recurrent neural networks (RNNs), which achieve multi-step forecasts through recursion. While these methods are effective in some applications, achieving high forecast accuracy often suffers from long training times and accumulated errors. Therefore, a deep learning forecasting algorithm for industrial time series data based on trend and residual separation is needed to address these issues. Summary of the Invention
[0004] The technical problem to be solved by the present invention is to provide a deep learning industrial time series data prediction algorithm based on trend and residual separation, aiming to solve the problem that the existing technology relies on recursion for multi-step prediction, has a long training time and cumulative errors, resulting in poor accuracy. It has the characteristics of being able to perform parallel computing, significantly shortening the training time, and effectively avoiding the error accumulation problem in traditional recursive methods. It is particularly suitable for large-scale time series prediction tasks.
[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: The deep learning industrial time series data prediction algorithm based on trend and residual separation includes the following steps: S1, input historical time series data and control quantity data, and decompose the time series data and control quantity data into seasonal components and trend components respectively through moving average and trend decomposition; S2, extracts features and reduces the dimensionality of the decomposed seasonal and trend components. The seasonal and trend components of the control quantity are reduced to a one-dimensional representation through a linear mapping layer, and then stacked with the corresponding parts of the target sequence to generate the final seasonal and trend features. S3, seasonality and trend features are input into the fully connected layer and nonlinear activation function respectively to generate hidden representations, and then output independent prediction results after linear transformation. and , add up to get the complete prediction value ; S4 uses the back propagation algorithm to train the model with the goal of minimizing the mean square error between the predicted value and the true value.
[0006] Preferably, in step S1, the historical time series data is defined as , the control quantity data is defined as and ,in represents the observed value of the time series at time t, Indicates the historical and future parts of the control quantity data, represents the number of time steps, represents the time step index, and L represents the time step length.
[0007] Preferably, the decomposition operation in step S1 specifically includes: For time series data Calculate seasonal components via moving average filters and trend components : ; ; ; Control quantity data Perform the same operation to obtain the seasonal component and trend components ; is the window size, which indicates the calculation window width of the moving average filter, i.e., k data points; Indicates that in calculation , iterates over the loop variables at the time points.
[0008] Preferably, the feature extraction and dimensionality reduction in step S2 includes: Controlling seasonal components and trend components Through the linear mapping matrix 、 Reduce the dimension to one dimension and get a one-dimensional representation and .
[0009] Preferably, feature extraction and dimensionality reduction further include: The reduced dimension features are stacked with the corresponding parts of the target sequence to generate the final features and : ; ; Where, and Represents the linear mapping matrix 、 The linear transformation matrix of .
[0010] Preferably, in the feature fusion and prediction of step S3, seasonality and trend features are processed respectively by: Seasonal characteristics Generate hidden representation through fully connected layers and nonlinear activation , and then linearly transform to output a separate seasonal forecast ; Trend characteristics Generate hidden representation through fully connected layers and nonlinear activation , and then output a separate trend forecast through linear transformation .
[0011] Preferably, the seasonal and trend forecasts are summed to get the complete forecast: ; Where, and There are separate seasonal forecasts and separate trend forecasts.
[0012] Preferably, in step S4, the training process adopts the mean square error loss function: ; in is the number of samples, is the predicted value, is the true value.
[0013] Preferably, the deep learning industrial time series data prediction device based on trend and residual separation includes: Data decomposition module: used to receive historical time series data and control quantity data, and decompose the input data into seasonal components and trend components through the moving average filter and trend decomposition unit; A feature processing module, connected to the data decomposition module, includes: The linear mapping unit is used to transform the seasonal component and trend component of the control quantity into the matrix 、 Dimensionality reduction to one-dimensional representation; Feature stacking unit is used to stack the reduced dimension features with the corresponding parts of the target sequence and pass the linear transformation matrix 、 Generate final seasonality and trend characteristics; The prediction module is connected to the feature processing module and includes: Seasonal prediction unit, which contains fully connected layers and nonlinear activation functions to process seasonal features and output individual seasonal predictions; The trend forecast unit has the same structure as the seasonal forecast unit and is used to process trend features and output a separate trend forecast; Adder, used to perform Get the final predicted value; a training optimization module, connected to the prediction module, configured to adjust model parameters by a back propagation algorithm with the goal of minimizing the mean square error between the predicted value and the true value; The device is used to implement the above-mentioned deep learning industrial time series data prediction algorithm based on trend and residual separation.
[0014] An electronic device includes a processor and a memory, wherein the memory stores a computer program. When the computer program is executed by the processor, the deep learning industrial time series data prediction algorithm based on trend and residual separation is implemented.
[0015] The beneficial effects of the present invention are as follows: 1. Parallel computing avoids recursive error accumulation: Traditional time series prediction algorithms with control variables are typically based on recurrent neural networks, which implement multi-step predictions through recursion. However, this approach is prone to cumulative errors in long-term time series predictions and requires long training times. This algorithm uses parallel computing to predict future time series, avoiding the recursive process and thus eliminating the problem of error accumulation.
[0016] 2. Independent Processing of Trend and Seasonal Components: This algorithm decomposes the input time series data into trend and seasonal components, modeling and forecasting the future values of each component separately. This decomposition strategy enables the model to more effectively capture both long-term trends and short-term fluctuations in time series data, significantly improving forecast accuracy. Unlike traditional methods that treat time series data as a single input, this method has a more targeted structure and can more flexibly adapt to changes of varying frequencies, thereby enhancing the model's adaptability to complex time series patterns.
[0017] 3. Incorporating Future Control Variable Information: This algorithm not only uses historical control variable data but also incorporates prior information about future control variables when predicting future target sequences. This innovative design effectively improves the accuracy of future time series predictions and is particularly suitable for industrial applications with external control factors, such as hydropower generation and energy load forecasting. The inclusion of future control variables better reflects actual operating conditions, making the model more adaptable to various environments and scenarios.
[0018] 4. Simplified Structure and Improved Computational Efficiency: Through linear layer dimensionality reduction and feature stacking, this algorithm simplifies the model structure and reduces computational complexity while maintaining prediction accuracy. Compared to traditional recursive methods, this algorithm's parallel computing process significantly reduces training and inference time, making it more suitable for real-time prediction needs in industrial environments. Furthermore, feature dimensionality reduction reduces computing resource usage, significantly improving model efficiency and making it suitable for deployment on resource-constrained devices. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 It is a schematic diagram of the overall process of the system of the present invention; Figure 2 It is a connection block diagram of the device part of the present invention; Figure 3 3 is a schematic diagram of a comparison curve diagram of an embodiment of the present invention and an RNN algorithm. DETAILED DESCRIPTION
[0020] Example 1: like Figure 1 As shown in the figure, the deep learning industrial time series data prediction algorithm based on trend and residual separation includes the following steps: S1, input historical time series data and control quantity data, and decompose the time series data and control quantity data into seasonal components and trend components respectively through moving average and trend decomposition; S2, extracts features and reduces the dimensionality of the decomposed seasonal and trend components. The seasonal and trend components of the control quantity are reduced to a one-dimensional representation through a linear mapping layer, and then stacked with the corresponding parts of the target sequence to generate the final seasonal and trend features. S3, seasonality and trend features are input into the fully connected layer and nonlinear activation function respectively to generate hidden representations, and then output independent prediction results after linear transformation. and , add up to get the complete prediction value ; S4 uses the back propagation algorithm to train the model with the goal of minimizing the mean square error between the predicted value and the true value.
[0021] Preferably, in step S1, the historical time series data is defined as , the control quantity data is defined as and ,in represents the observed value of the time series at time t, Indicates the historical and future parts of the control quantity data, represents the number of time steps, represents the time step index, and L represents the time step length.
[0022] Preferably, the decomposition operation in step S1 specifically includes: For time series data Calculate seasonal components via moving average filters and trend components : ; ; ; Control quantity data Perform the same operation to obtain the seasonal component and trend components ; is the window size, which indicates the calculation window width of the moving average filter, i.e., k data points; Indicates that in calculation , iterates over the loop variables at the time points.
[0023] Preferably, the feature extraction and dimensionality reduction in step S2 includes: Controlling seasonal components and trend components Through the linear mapping matrix 、 Reduce the dimension to one dimension and get a one-dimensional representation and .
[0024] Preferably, feature extraction and dimensionality reduction further include: The reduced dimension features are stacked with the corresponding parts of the target sequence to generate the final features and : ; ; Where, and Represents the linear mapping matrix 、 The linear transformation matrix of .
[0025] Preferably, in the feature fusion and prediction of step S3, seasonality and trend features are processed respectively by: Seasonal characteristics Generate hidden representation through fully connected layers and nonlinear activation , and then linearly transform to output a separate seasonal forecast ; Trend characteristics Generate hidden representation through fully connected layers and nonlinear activation , and then output a separate trend forecast through linear transformation .
[0026] Preferably, the seasonal and trend forecasts are summed to get the complete forecast: ; Where, and There are separate seasonal forecasts and separate trend forecasts.
[0027] Preferably, in step S4, the training process adopts the mean square error loss function: ; in is the number of samples, is the predicted value, is the true value.
[0028] like Figure 2 As shown, preferably, the deep learning industrial time series data prediction device based on trend and residual separation includes: Data decomposition module: used to receive historical time series data and control quantity data, and decompose the input data into seasonal components and trend components through the moving average filter and trend decomposition unit; A feature processing module, connected to the data decomposition module, includes: The linear mapping unit is used to transform the seasonal component and trend component of the control quantity into the matrix 、 Dimensionality reduction to one-dimensional representation; Feature stacking unit is used to stack the reduced dimension features with the corresponding parts of the target sequence and pass the linear transformation matrix 、 Generate final seasonality and trend characteristics; The prediction module is connected to the feature processing module and includes: Seasonal prediction unit, which contains fully connected layers and nonlinear activation functions to process seasonal features and output individual seasonal predictions; The trend forecast unit has the same structure as the seasonal forecast unit and is used to process trend features and output a separate trend forecast; Adder, used to perform Get the final predicted value; a training optimization module, connected to the prediction module, configured to adjust model parameters by a back propagation algorithm with the goal of minimizing the mean square error between the predicted value and the true value; The device is used to implement the above-mentioned deep learning industrial time series data prediction algorithm based on trend and residual separation.
[0029] An electronic device includes a processor and a memory, wherein the memory stores a computer program. When the computer program is executed by the processor, the deep learning industrial time series data prediction algorithm based on trend and residual separation is implemented.
[0030] Example 2: The following is an implementation example designed based on the patent solution, constructing a hydropower application scenario strictly based on the technical briefing document, and comparing it with the traditional RNN method: The deep learning industrial time series data prediction algorithm based on trend and residual separation disclosed in the present invention is applied to a hydropower plant power generation load forecasting project in a prefecture-level city in central China: 1. The experimental settings are shown in Table 1 below: Table 1: Experimental setup.
[0031] 2. The performance comparison results are shown in Table 2 below: Table 2: Performance comparison results
[0032] 3. Principle of advanced realization: (1) Error accumulation control: The traditional RNN recursive prediction shows significant error diffusion after 48 hours (for example, the valve opening fluctuation is amplified step by step). However, this method uses a parallel decomposition-prediction architecture: The trend component directly studies the long-term changes in water levels; The seasonal component captures short-term fluctuations in flow; There is no error transfer path when adders are fused; (2) Future control volume value: Introducing weather forecast u_future The prediction accuracy is critically improved, as shown in Table 3 below: Table 3: Introduction of weather forecast u_future Comparison:
[0033] It can be seen from Table 3 that the future control quantity plays a decisive role in the prediction of sudden operating conditions of hydropower generation.
[0034] (3) Computational efficiency advantage: Dimensionality reduction and simplification: The control volume feature is reduced from the original 128 dimensions to 1 dimension, which significantly reduces the computational complexity of the fully connected layer; Parallel processing: trend / seasonal branches can be run simultaneously; Resource optimization: The reduced number of parameters makes embedded deployment possible.
[0035] Typical prediction results are as follows Figure 3 As shown, Figure 3 It is shown in the figure that RNN overestimates the peak power generation during the rainstorm period due to cumulative errors, while this method accurately captures the stable output after reservoir regulation.
Claims
1. A deep learning industrial time series data prediction algorithm based on trend and residual separation, characterized by: The following steps are involved: S1, input historical time series data and control quantity data, and decompose the time series data and control quantity data into seasonal components and trend components respectively through moving average and trend decomposition; S2, extracts features and reduces the dimensionality of the decomposed seasonal and trend components. The seasonal and trend components of the control quantity are reduced to a one-dimensional representation through a linear mapping layer, and then stacked with the corresponding parts of the target sequence to generate the final seasonal and trend features. S3, seasonality and trend features are input into the fully connected layer and nonlinear activation function respectively to generate hidden representations, and then output independent prediction results after linear transformation. and , add up to get the complete prediction value ; S4 uses the back propagation algorithm to train the model with the goal of minimizing the mean square error between the predicted value and the true value.
2. The deep learning industrial time series data prediction algorithm based on trend and residual separation according to claim 1 is characterized in that: In step S1, the historical time series data is defined as , the control quantity data is defined as and ,in represents the observed value of the time series at time t, Represents the historical and future parts of the control quantity data; represents the number of time steps, represents the time step index, and L represents the time step length.
3. The deep learning industrial time series data prediction algorithm based on trend and residual separation according to claim 2 is characterized in that: The decomposition operation in step S1 specifically includes: For time series data Calculate seasonal components via moving average filters and trend components : ； ； ； Control quantity data Perform the same operation to obtain the seasonal component and trend components ; is the window size, which indicates the calculation window width of the moving average filter, i.e., k data points; Indicates that in calculation , iterates over the loop variables at the time points.
4. The deep learning industrial time series data prediction algorithm based on trend and residual separation according to claim 3 is characterized in that: The feature extraction and dimensionality reduction in step S2 include: Controlling seasonal components and trend components Through the linear mapping matrix 、 Reduce the dimension to one dimension and get a one-dimensional representation and .
5. The deep learning industrial time series data prediction algorithm based on trend and residual separation according to claim 4 is characterized in that: Feature extraction and dimensionality reduction also include: The reduced dimension features are stacked with the corresponding parts of the target sequence to generate the final features and : ； ； Where, and Represents the linear mapping matrix 、 The linear transformation matrix of .
6. The deep learning industrial time series data prediction algorithm based on trend and residual separation according to claim 5 is characterized in that: In the feature fusion and prediction of step S3, seasonality and trend features are processed as follows: Seasonal characteristics Generate hidden representation through fully connected layers and nonlinear activation , and then linearly transform to output a separate seasonal forecast ; Trend characteristics Generate hidden representation through fully connected layers and nonlinear activation , and then output a separate trend forecast through linear transformation .
7. The deep learning industrial time series data prediction algorithm based on trend and residual separation according to claim 6 is characterized in that: Add the seasonal and trend forecasts together to get the complete forecast: ； Where, and There are separate seasonal forecasts and separate trend forecasts.
8. The deep learning industrial time series data prediction algorithm based on trend and residual separation according to claim 1 is characterized in that: In step S4, the training process uses the mean square error loss function: ； in is the number of samples, is the predicted value, is the true value.
9. A deep learning industrial time series data prediction device based on trend and residual separation, characterized in that: The device includes: Data decomposition module: used to receive historical time series data and control quantity data, and decompose the input data into seasonal components and trend components through the moving average filter and trend decomposition unit; A feature processing module, connected to the data decomposition module, includes: The linear mapping unit is used to transform the seasonal component and trend component of the control quantity into the matrix 、 Dimensionality reduction to one-dimensional representation; Feature stacking unit is used to stack the reduced dimension features with the corresponding parts of the target sequence and pass the linear transformation matrix 、 Generate final seasonality and trend characteristics; The prediction module is connected to the feature processing module and includes: Seasonal prediction unit, which contains fully connected layers and nonlinear activation functions to process seasonal features and output individual seasonal predictions; The trend forecast unit has the same structure as the seasonal forecast unit and is used to process trend features and output a separate trend forecast; Adder, used to perform Get the final predicted value; a training optimization module, connected to the prediction module, configured to adjust model parameters by a back propagation algorithm with the goal of minimizing the mean square error between the predicted value and the true value; The device is used to implement the deep learning industrial time series data prediction algorithm based on trend and residual separation as described in any one of claims 1-8.
10. An electronic device, characterized in that: The invention comprises a processor and a memory, wherein the memory stores a computer program. When the computer program is executed by the processor, the deep learning industrial time series data prediction algorithm based on trend and residual separation as described in any one of claims 1 to 8 is implemented.