Depth time sequence prediction method and system based on event gating mechanism

Abstract

The invention discloses a depth time sequence prediction method and system based on an event gating mechanism, and belongs to the technical field of artificial intelligence, and the method comprises the steps: obtaining time sequence data and corresponding event sequence data; mapping the time sequence data and the corresponding event sequence data to a low-dimensional potential vector space to obtain a time sequence feature representation and an event sequence feature representation; according to the time step sequence, inputting the time sequence feature representation and the event sequence feature representation into an improved recurrent neural network for processing to obtain a hidden state sequence; performing attention calculation based on the hidden state sequence to obtain final vector representation; and obtaining a prediction result according to the final vector representation. According to the invention, the prediction capability of the time sequence containing the non-periodic event interference can be improved.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to a deep time series prediction method and system based on an event gating mechanism. Background Art

[0002] The modeling and prediction of time series data are of great significance in many fields such as logistics scheduling, equipment failure warning, and medical monitoring. Many real-world time series contain both periodic patterns (such as daily or seasonal trends) and are disturbed by non-periodic events. Such non-periodic events (such as natural disasters, policy changes, sudden equipment failures, medical emergencies, etc.) are often unpredictable and can cause abnormal fluctuations in time series or deviations from normal trajectories. This poses a great challenge to traditional forecasting methods: the model must not only capture long-term periodic trends, but also respond to the impact of emergencies in a timely manner. In recent years, with the development of deep learning, a variety of models and methods for time series forecasting have emerged. For example:

[0003] ① Transformer model: The Transformer is a deep learning model based on the self-attention mechanism and has been widely used in tasks such as natural language processing and time series prediction. Compared with traditional recurrent neural networks, the Transformer has strong parallel computing capabilities and is good at capturing long-range dependencies.

[0004] ②Pyraformer model: Pyraformer is an improved Transformer model based on the low-complexity pyramidal attention mechanism, primarily used for efficient modeling and prediction of long-series temporal data. (file-vdpc1yztpjhpgln7q95cvg) This model reduces the computational overhead of long-series modeling through hierarchical attention calculations, improving long-series prediction performance to a certain extent.

[0005] ③ Non-Stationary Transformer (NSTransformer-S): NSTransformer-S is a simplified version of the NSTransformer model, designed to address the stabilization issues encountered in other methods. By introducing sequence stabilization and destabilization attention mechanisms, NSTransformer-S makes the model more stable in improving prediction performance.

[0006] The above methods have made significant progress in capturing long-range dependencies and improving forecasting accuracy. However, these existing models generally assume that time series patterns are relatively stable and lack specialized mechanisms for handling non-periodic, unexpected events that occur in the series. When external events cause abnormal fluctuations in the data, traditional models tend to either ignore them as noise or simply incorporate the event information directly into the model input. The former may miss key information, while the latter, due to the lack of targeted structural support, cannot fully utilize the role of event information. Therefore, in complex time series prediction tasks involving unexpected event interference, existing models still have significant shortcomings. It is necessary to design new time series modeling methods to more effectively integrate event information and improve robustness to event disturbances.

[0007] In summary, the prior art still has the following technical defects.

[0008] ① Lack of event fusion mechanisms: Commonly used recurrent neural networks (such as LSTM) and other forecasting models lack specialized structures or gating mechanisms for non-periodic events. This results in models being unable to fully integrate event information with the primary time series data when processing sequences containing unexpected events. For example, when unexpected orders occur in logistics forecasts or urgent failures occur in equipment monitoring, traditional models often struggle to specifically incorporate the impact of these events on subsequent trends, resulting in an incomplete understanding of the sequence pattern.

[0009] ② Insufficient noise immunity: Existing models are susceptible to interference from unusual fluctuations and noisy data, making it difficult to focus on the impact of key events on long-term trends. When a time series is interspersed with a large amount of noise or short-term fluctuations unrelated to the long-term trend, the model may mistakenly interpret these as part of the trend, thus deviating from the correct forecast direction. For example, random fluctuations in device sensor data and one-time measurement errors in medical monitoring can interfere with the model's judgment of the overall trend. Existing methods lack effective mechanisms to distinguish truly important event signals from background noise, resulting in reduced forecasting performance. Summary of the Invention

[0010] To address the above problems, the present invention proposes a deep time series prediction method and system based on an event gating mechanism. The core of the method is to fuse event information on the basis of a long short-term memory network (LSTM) and combine it with an attention mechanism to improve the prediction ability of time series with non-periodic event interference.

[0011] To achieve the above objectives, the technical solution of the present invention includes the following contents.

[0012] A deep time series prediction method based on an event gating mechanism, the method comprising:

[0013] Obtain time series data and corresponding event series data;

[0014] Map the time series data and the corresponding event sequence data into a low-dimensional latent vector space to obtain the time series feature representation and the event sequence feature representation;

[0015] According to the time step sequence, the time series feature representation and the event sequence feature representation are input into an improved recurrent neural network for processing to obtain a hidden state sequence; wherein the improved recurrent neural network has a two-layer structure;

[0016] Perform attention calculation based on the hidden state sequence to obtain a final vector representation;

[0017] The prediction result is obtained based on the final vector representation.

[0018] Furthermore, obtaining time series data and corresponding event series data includes:

[0019] Obtain original time series data and corresponding original event series data;

[0020] The original time series data and the corresponding original event series data are normalized and aligned.

[0021] Furthermore, the time series data and the corresponding event sequence data are mapped to a low-dimensional latent vector space to obtain the time series feature representation and the event sequence feature representation, including:

[0022] Each dimension in the time series data is input into the feature p t Encoded into a fixed-dimensional embedding vector X through a linear layer or a multi-layer perceptron t ;

[0023] Each dimension in the time series data is input into feature a t Encoded into a fixed-dimensional embedding vector N via a linear layer, multi-layer perceptron encoding, or an embedding dictionary encoding combined with event types t ;

[0024] Unified embedding vector X t and embedding vector N t format.

[0025] Furthermore, the first layer of the improved recurrent neural network is composed of a plurality of EventLSTMA units, and the second layer of the improved LSTM network layer is composed of a plurality of standard LSTM units; wherein, the EventLSTMA unit adds an event control gate on the basis of the standard LSTM unit, and embeds the event vector N based on the event. t and the hidden state h at the t-1th time step t-1 To generate the gate variable of the event control gate to expand the input gate of the standard LSTM unit, the event embedding vector Nt is the event feature representation of the tth time step in the event sequence feature representation.

[0026] Furthermore, the processing of the EventLSTMA unit includes:

[0027] Get the time embedding vector X of the tth time step in the time series feature representation t , event sequence feature represents the event embedding vector N at the tth time step in the event t , the hidden state h at the t-1th time step 0,t-1 And the memory state C at the t-1th time step t-1 ;

[0028] Based on the time embedding vector X t and hidden state h 0,t-1 , respectively calculate the forget gate f t , input gate i t , output gate o t and candidate memory states

[0029] Based on the event embedding vector N t and hidden state h 0,t-1 , calculate event control gate n t ;

[0030] Generate the memory state of the tth time step

[0031] According to the memory state C t and the gate variable o of the output gate t , get the hidden state h at the tth time step 0,t .

[0032] Furthermore, the first layer of the improved recurrent neural network is composed of several EventLSTMB units, and the second layer of the improved LSTM network layer is composed of several standard LSTM units; wherein, the EventLSTMB unit introduces an independent event gating channel based on the standard LSTM unit, and is based on the event embedding vector N t and the hidden state h at the t-1th time step t-1 To generate a gate variable n t , to regulate the event control gate in the event-gated channel, the event embedding vector N t is the event feature representation of the tth time step in the event sequence feature representation.

[0033] Furthermore, the processing of the EventLSTMB unit includes:

[0034] Get the time embedding vector X of the tth time step in the time series feature representation t , event sequence feature represents the event embedding vector N at the tth time step in the event t , the hidden state h at the t-1th time step 0,t-1 And the memory state C at the t-1th time step t-1 ;

[0035] Based on the time embedding vector X t and hidden state h 0,t-1 , respectively calculate the forget gate f t , input gate i t , output gate o t and candidate memory states

[0036] Based on the event embedding vector N t and hidden state h 0,t-1 , respectively calculate the event control gate n t and candidate memory states

[0037] Generate the memory state of the tth time step

[0038] According to the memory state C t and output gate o t , get the hidden state h at the tth time step 0,t .

[0039] Furthermore, according to the time step sequence, the time series feature representation and the event sequence feature representation are input into an improved recurrent neural network for processing to obtain a hidden state sequence, including:

[0040] The time series feature representation and the event sequence feature representation are input into the first layer of the improved LSTM network layer to obtain the hidden state sequence H0=h 0,1 ,h 0,2 ,…,h 0,t ;

[0041] Set the hidden state sequence H0 = h 0,1 ,h 0,2 ,…,h 0,t Input to the second layer of the improved LSTM network layer to obtain the hidden state sequence H1=h 1,1 ,h 1,2 ,…,h 1,t .

[0042] Furthermore, attention calculation is performed based on the hidden state sequence to obtain a final vector representation, including:

[0043] Using an attention mechanism to assign a weight to each hidden state in the hidden state sequence, the weight is used to measure the contribution of each time step to the prediction target;

[0044] The hidden states are weighted based on the weights to obtain the final vector representation.

[0045] A deep time series prediction system based on an event gating mechanism, the system comprising:

[0046] Input module, used to obtain time series data and corresponding event sequence data;

[0047] Embedding module, which is used to map time series data and corresponding event sequence data into a low-dimensional latent vector space to obtain time series feature representation and event sequence feature representation;

[0048] A hidden state calculation module is used to input the time series feature representation and the event sequence feature representation into an improved recurrent neural network for processing according to the time step sequence to obtain a hidden state sequence; wherein the improved recurrent neural network has a two-layer structure;

[0049] An attention module, configured to perform attention calculation based on the hidden state sequence to obtain a final vector representation;

[0050] The prediction module is used to obtain a prediction result based on the final vector representation.

[0051] Compared with the prior art, the present invention has at least the following beneficial effects.

[0052] This paper introduces a specialized event-control gate unit into the LSTM architecture, implicitly fusing aperiodic event information with periodic sequence data during cell state updates. This mechanism allows the model to flexibly incorporate unexpected event factors and dynamically adjust its response to events during internal state updates, improving its adaptability to unconventional information.

[0053] This paper uses a layered LSTM network to process time series data in two stages: the bottom layer focuses on event information fusion, while the upper layer deeply extracts long-term dependency features. This structure improves the model's expressiveness and ability to capture complex patterns, allowing it to focus on local event impacts while retaining the overall trend context.

[0054] This paper introduces an attention mechanism to weight historical hidden states, highlighting the contribution of key moments to prediction. By assigning higher weight to useful information and reducing noise interference, the model can still focus on truly important signals in environments containing noise and abnormal fluctuations, thereby improving prediction accuracy and stability.

[0055] Based on the above technical features, the present invention has good domain versatility and anti-disturbance characteristics. Specifically, the model structure has no special dependence on the data domain and can be widely applied to various time series tasks with non-periodic event disturbances, such as logistics demand forecasting, industrial fault warning, and medical indicator monitoring. Compared with models designed only for a certain field, this method demonstrates robust modeling capabilities on different types of data sets. When sudden events occur in the sequence, the model can maintain strong robustness and accurately capture the impact of the event on the sequence without excessively deviating from the long-term trend. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Flowchart of the deep temporal modeling method based on event gating mechanism.

[0057] Figure 2 Schematic diagram of the structure of EventLSTMA cells.

[0058] Figure 3 Schematic diagram of the structure of EventLSTMB cells.

[0059] Figure 4 Comparison of the mean values ​​of evaluation indicators of the ablation experiment model. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below through specific implementation and in conjunction with the accompanying drawings.

[0061] The overall architecture of the model of the present invention is as follows: Figure 1As shown in the figure, the system comprises an input layer, an embedding layer, an LSTM layer, an attention layer, and a prediction layer. First, the historical time series data and the corresponding event sequence are input into the embedding layer, where they are mapped into a low-dimensional latent vector space, respectively, to obtain time series feature representations and event feature representations. These representations are then sequentially fed into an improved LSTM network layer for processing. An event-gating mechanism is introduced in the LSTM layer to simultaneously learn the temporal dependencies of the historical sequence and the influence of non-periodic events on the sequence. This LSTM layer employs a two-layer architecture: the first layer comprises an improved event-gated LSTM unit, which locally fuses event and time series information; the second layer comprises traditional LSTM units, which extract higher-level temporal features and long-term dependencies. Next, an attention layer is placed above the LSTM layer to dynamically model the contribution of the hidden state at each time step to the prediction target, highlighting information related to key events and mitigating the influence of noise. Finally, the prediction layer feeds the fused historical information (weighted by attention) into a fully connected network, outputting a prediction for the next time step. The predicted value can be the target variable value at the next moment, such as demand for the next period, sensor readings, etc. The entire model architecture is applicable to all types of time series forecasting tasks involving event disturbances and is not limited to a single application domain.

[0062] 1. Input layer.

[0063] The input data entering the input layer includes two categories:

[0064] 1. Periodic main sequence data P = {p1, p2, ..., p t}, where p t Represents the original multidimensional features at the tth time step (such as sensor readings, indicator data, operation records, etc.);

[0065] 2. Non-periodic event A={a1,a2,…,a t}, where a t Represents the event feature vector aligned with the tth time step. If there is no event at a time step, let a t =0.

[0066] In one embodiment, data needs to be normalized and aligned before entering the input layer. Normalization applies min-max normalization to the numerical features of each dimension in the periodic input, compressing them to the interval [0, 1]. Event alignment is performed because events do not have a fixed sampling frequency. Events with irregular timestamps need to be aligned to the corresponding time step by weighted aggregation of events within the window or filling in a default vector form.

[0067] 2. Embedding Layer

[0068] As the first processing module, the embedding layer is responsible for uniformly encoding the input data. The encoding process specifically includes the following steps.

[0069] Step 2.1: Embed the input data to obtain the embedding vector.

[0070] For the main sequence data P, each dimension input feature p t Encoded into a fixed-dimensional embedding vector via a linear layer (or multi-layer perceptron)

[0071] For non-periodic events A, they are encoded into event embedding vectors in the same way (or combined with the embedding dictionary of event types)

[0072] Step 2.2: Unify the embedding vector and embedding vector The input format of .

[0073] The embedding vector of the unified input format provides a low-dimensional dense representation for the subsequent event gate control mechanism and LSTM processing. At each time step t, the embedding layer outputs the unified time series embedding vector and event embedding vector As the input of the subsequent LSTM layer.

[0074] 3. LSTM layer.

[0075] This paper designs two improved event-gated LSTM cell structures in the LSTM layer, denoted as EventLSTMA and EventLSTMB, respectively, to explicitly integrate event information within the LSTM:

[0076] EventLSTMA: This unit adds an "event control gate" to the traditional LSTM. Specifically, it combines the current event input with the main time series input to form a new gating variable. This event control gate uses non-periodic event data as a modulating factor, expanding the input gate of the original LSTM, allowing the current unit's state update to account for the additional impact of the event. Through this mechanism, EventLSTMA combines historical sequence data with current event information to generate new candidate inputs. When an event occurs, the corresponding gating variable assigns greater weight to the event, thereby emphasizing the critical role of non-periodic events in the current state during cell state updates. This design effectively ensures that the model can promptly integrate event information and adjust its state at the cellular level when encountering unexpected events.

[0077] The formula is as follows:

[0078] n t =σ(W n ·[h t-1 ,N t ]+b n )

[0079]

[0080] Among them, n t is a new gating variable that combines input and event information, W n and b n are the relevant weight matrix and bias term, h t-1 represents the hidden state of the previous time step, f t It is the forget gate, i t is the input gate, is the memory state of the current time step, is a candidate memory state.

[0081] EventLSTMB: This cell architecture introduces an independent event-gated channel within the LSTM. Unlike EventLSTMA, where the event gate is coupled to the input gate, EventLSTMB provides an independent processing path for non-periodic event information. The added event gate does not interfere with the existing input, forget, and output gates, but rather dynamically models event features in parallel. Specifically, EventLSTMB computes two candidate cell states at each time step: one derived from the periodic main sequence data (generated by the original LSTM gating) and the other derived specifically from non-periodic event data (generated by the event gate). These two candidate states are then fused to update the final cell state for that time step. This means that whenever an event occurs, the event information is independently gated to generate a separate memory representation, which is then merged with the regular sequence memory. This design ensures a more flexible model response to unexpected events: event information has an independent processing path within the model and is not overwhelmed by periodic patterns, while the periodic patterns of the main sequence are not excessively forgotten due to the addition of events. Ultimately, EventLSTMB enables the model to balance the impact of periodic trends and non-periodic events when updating the state, giving the two equal status in modeling.

[0082] The formula is as follows:

[0083] n t =σ(W n ·[h t-1 ,N t ]+b n )

[0084]

[0085] in, and Represent two different candidate memory states, Calculated by the current input and the hidden state of the previous time step, the two candidate memory states work together to affect the final memory state

[0086] In one embodiment, to further enhance the model's ability to depict temporal relationships and improve prediction accuracy, the present invention adopts a two-layer LSTM recurrent network structure. The first layer uses the above-mentioned improved event LSTM unit (whether EventLSTMA or EventLSTMB is selected), which is mainly used to fully explore and fuse the correlation pattern between event information and the underlying time series at the bottom layer; the second layer uses a standard LSTM unit and accepts the output sequence of the first layer as input. Through the fusion of events and sequences in the first layer, the model obtains a feature sequence that includes the impact of events; the second layer continues to explore more complex temporal features on this basis, capturing long-term dependencies and global trends. The multi-level recurrent structure enables the model to extract features from fine-grained to macroscopic levels layer by layer: the lower layer highlights the effect of events on short-term changes, and the upper layer focuses on long-term evolution and trend smoothing, thereby improving the overall resolution and prediction stability of complex temporal patterns. This hierarchical design enables the model to adapt to short-term sharp fluctuations while characterizing long-term trends, and exhibits higher flexibility and robustness when facing non-periodic emergencies.

[0087] In another embodiment, in addition to the above LSTM-based implementation, there are some variations that can achieve similar invention objectives. For example, the event gating mechanism can be applied to other recurrent neural units (such as GRU gated recurrent units) or time series models to replace LSTM as the basic sequence model.

[0088] 4. Attention layer.

[0089] To enable the model to more fully utilize historical information and focus on key event features, the present invention introduces an attention mechanism on top of the two-layer LSTM. When processing long sequences, traditional LSTMs often rely solely on the hidden state at the final moment for prediction, which may cause some important historical information to be ignored, especially when this information is relevant to the prediction target but occurs at earlier time steps. To this end, the present invention adds a layer of attention network, taking the hidden state of the two-layer LSTM at each time step as input. By calculating attention weights with the hidden state at the final moment (or prediction query vector), it measures the contribution of each time step to the prediction target. Specifically, the attention mechanism assigns a weight to each hidden state, the size of which reflects the importance of the information at that moment to the current prediction. Subsequently, all hidden states are weighted and summed according to their respective weights to obtain a comprehensive representation that incorporates key moment information. Through this attention weighting, the model can highlight the influence of key moments related to non-periodic events (such as the moment when important events occur) during the prediction phase, while minimizing the impact of irrelevant noise fluctuations. The introduction of the attention mechanism effectively improves the model's ability to extract long-range dependencies and key event information, making the prediction results more accurate and reliable.

[0090] 5. Prediction layer.

[0091] The fused historical information (weighted by attention) is fed into a fully connected network, which outputs a predicted value for the next time step. This predicted value can be the value of the target variable at the next moment, such as the demand for the next period, sensor readings, etc.

[0092] The following experiment further illustrates the deep temporal modeling method based on the event gating mechanism of the present invention.

[0093] To demonstrate the significant effectiveness of this invention in integrating event information for time series forecasting, this experiment was conducted on multiple real-world datasets. The results show that the proposed method outperforms existing typical models in both prediction accuracy and robustness. In particular, in scenarios with non-periodic event interference, the proposed model demonstrates lower prediction error and more stable performance compared to traditional baseline methods such as LSTM and Transformer.

[0094] Specifically, in terms of evaluation indicators, the experiment used indicators such as mean absolute error (MAE) and median absolute error (MedAE) to evaluate the prediction results. Compared with existing methods, such as Figure 4As shown in Table 1, the average values ​​of MAE and MedAE of the model of the present invention are significantly reduced, indicating that the prediction accuracy is improved. For example, compared with the standard baseline LSTM model, the error of the model is significantly reduced after the event gating mechanism is added; after further integrating the attention mechanism on this basis, the prediction error is further reduced, and the model achieves the best performance. This proves that the event gating unit and the attention module play a key role in improving the model's ability to handle non-periodic events. Combining various comparisons, the present method has achieved optimal or near-optimal results in most test scenarios, verifying the effectiveness of combining event information with LSTM and enhancing time series feature extraction through the attention mechanism. Compared with the existing technology, the present invention not only improves the prediction accuracy of event-perturbed sequences, but also significantly improves the robustness of the model under noise interference, and can be more reliably applied in actual complex environments.

[0095] Table 1 Comparison of the mean values ​​of the overall evaluation indicators of each model

[0096]

[0097] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the present disclosure. The present disclosure is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be considered merely as exemplary, and the present disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and variations may be made without departing from the scope thereof.

Claims

1. A deep time series prediction method based on event gating mechanism, characterized in that: The method comprises: Obtain time series data and corresponding event series data; Map the time series data and the corresponding event sequence data into a low-dimensional latent vector space to obtain the time series feature representation and the event sequence feature representation; According to the time step sequence, the time series feature representation and the event sequence feature representation are input into an improved recurrent neural network for processing to obtain a hidden state sequence; wherein the improved recurrent neural network has a two-layer structure; Perform attention calculation based on the hidden state sequence to obtain a final vector representation; The prediction result is obtained based on the final vector representation.

2. The method according to claim 1, characterized in that Obtain time series data and corresponding event series data, including: Obtain original time series data and corresponding original event series data; The original time series data and the corresponding original event series data are normalized and aligned.

3. The method according to claim 1, characterized in that Map the time series data and the corresponding event sequence data into a low-dimensional latent vector space to obtain the time series feature representation and event sequence feature representation, including: Each dimension in the time series data is input into the feature p t Encoded into a fixed-dimensional embedding vector X through a linear layer or a multi-layer perceptron t ; Each dimension in the time series data is input into feature a t Encoded into a fixed-dimensional embedding vector N via a linear layer, multi-layer perceptron encoding, or an embedding dictionary encoding combined with event types t ; Unified embedding vector X t and embedding vector N t format.

4. The method according to claim 1, wherein The first layer of the improved recurrent neural network is composed of several EventLSTMA units, and the second layer of the improved LSTM network layer is composed of several standard LSTM units; wherein, the EventLSTMA unit adds an event control gate on the basis of the standard LSTM unit, and embeds the event vector N based on the event. t and the hidden state h at the t-1th time step t-1 To generate the gate variable of the event control gate to expand the input gate of the standard LSTM unit, the event embedding vector N t is the event feature representation of the tth time step in the event sequence feature representation.

5. The method according to claim 4, characterized in that The processing process of the EventLSTMA unit includes: Get the time embedding vector X of the tth time step in the time series feature representation t , event sequence feature represents the event embedding vector N at the tth time step in the event t , the hidden state h at the t-1th time step 0,t-1 And the memory state C at the t-1th time step t-1 ; Based on the time embedding vector X t and hidden state h 0,t-1 , respectively calculate the forget gate f t , input gate i t , output gate o t and candidate memory states Based on the event embedding vector N t and hidden state h 0,t-1 , calculate event control gate n t ; Generate the memory state of the tth time step According to the memory state C t and the gate variable o of the output gate t , get the hidden state h at the tth time step 0,t .

6. The method according to claim 1, characterized in that The first layer of the improved recurrent neural network is composed of several EventLSTMB units, and the second layer of the improved LSTM network layer is composed of several standard LSTM units; wherein, the EventLSTMB unit introduces an independent event gating channel based on the standard LSTM unit, and embeds the event vector N t and the hidden state h at the t-1th time step t-1 To generate a gate variable n t , to regulate the event control gate in the event-gated channel, the event embedding vector N t is the event feature representation of the tth time step in the event sequence feature representation.

7. The method according to claim 6, characterized in that The processing process of the EventLSTMB unit includes: Get the time embedding vector X of the tth time step in the time series feature representation t , event sequence feature represents the event embedding vector N at the tth time step in the event t , the hidden state h at the t-1th time step 0,t-1 And the memory state C at the t-1th time step t-1 ; Based on the time embedding vector X t and hidden state h 0,t-1 , respectively calculate the forget gate f t , input gate i t , output gate o t and candidate memory states Based on the event embedding vector N t and hidden state h 0,t-1 , respectively calculate the event control gate n t and candidate memory states Generate the memory state of the tth time step According to the memory state C t and output gate o t , get the hidden state h at the tth time step 0,t .

8. The method according to claim 5 or 7, characterized in that According to the time step sequence, the time series feature representation and event sequence feature representation are input into an improved recurrent neural network for processing to obtain a hidden state sequence, including: The time series feature representation and the event sequence feature representation are input into the first layer of the improved LSTM network layer to obtain the hidden state sequence H0=h 0,1 ,h 0,2 ,…,h 0,t ; The hidden state sequence H0 = h 0,1 ,h 0,2 ,…,h 0,t Input to the second layer of the improved LSTM network layer to obtain the hidden state sequence H1=h 1,1 ,h 1,2 ,…,h 1,t .

9. The method according to claim 1, characterized in that Attention calculation is performed based on the hidden state sequence to obtain the final vector representation, including: Using an attention mechanism to assign a weight to each hidden state in the hidden state sequence, the weight is used to measure the contribution of each time step to the prediction target; The hidden states are weighted based on the weights to obtain the final vector representation.

10. A deep time series prediction system based on event gating mechanism, characterized in that: The system comprises: Input module, used to obtain time series data and corresponding event sequence data; Embedding module, which is used to map time series data and corresponding event sequence data into a low-dimensional latent vector space to obtain time series feature representation and event sequence feature representation; A hidden state calculation module is used to input the time series feature representation and the event sequence feature representation into an improved recurrent neural network for processing according to the time step sequence to obtain a hidden state sequence; wherein the improved recurrent neural network has a two-layer structure; An attention module, configured to perform attention calculation based on the hidden state sequence to obtain a final vector representation; The prediction module is used to obtain a prediction result based on the final vector representation.

Images