Regional power marketing management information platform data service intelligent recommendation method based on artificial intelligence

Abstract

The invention relates to the technical field of data services, in particular to a regional power marketing management information platform data service intelligent recommendation method based on artificial intelligence, which comprises the following steps of: acquiring a user power consumption behavior sequence and an equipment alarm log from a power marketing platform, and performing event-driven timestamp resampling to obtain a user power consumption behavior sequence and an equipment alarm log; generating a multi-dimensional data stream with time synchronization; inputting the multi-dimensional data stream into a gated adversarial network, dynamically generating a mask matrix according to an equipment alarm level, stripping power consumption behavior distortion characteristics caused by equipment abnormity, and outputting a user reference demand vector; and performing similarity matching on the user reference demand vector and a power service knowledge base to generate a recommendation strategy set. According to the method, the stable and robust user reference demand vector is constructed, a data foundation is laid for subsequent recommendation strategy matching, and the technical bottleneck that the model of a traditional system is seriously distorted under abnormal fluctuation is broken through.

Description

Technical Field

[0001] The present invention relates to the field of data service technology, and in particular to an artificial intelligence-based intelligent recommendation method for data services of a regional power marketing management information platform. Background Art

[0002] With the gradual improvement of power market mechanisms and the diversification of user electricity demand, regional power marketing management is transitioning from the previous unified strategy issuance to a refined, intelligent service model. To improve user satisfaction and system operation and maintenance efficiency, power companies have generally built marketing management information platforms that include energy efficiency monitoring, user behavior analysis, and equipment operation and maintenance management. However, existing platforms still face the following technical bottlenecks in service strategy recommendation:

[0003] On the one hand, current policy recommendation systems are mostly based on rule matching or simple similarity calculation, lacking a multi-dimensional matching perspective. They are unable to simultaneously take into account key factors such as users' temporal behavior characteristics, sensitivity to policy costs, and device health status, resulting in a lack of targeted service strategies. In addition, recommendation weights are usually set statically and cannot be dynamically adjusted based on execution feedback. The system's adaptive capabilities are insufficient, making it difficult to cope with the policy adjustment needs brought about by changes in user needs and device status over time. Summary of the Invention

[0004] The present invention provides an artificial intelligence-based intelligent recommendation method for data services of a regional power marketing management information platform. The method integrates asynchronous multi-source data, has interference stripping capabilities, and supports multi-core similarity matching and adaptive weight optimization, so as to improve the technical level of the regional power marketing platform in terms of precise strategy matching and continuous recommendation optimization.

[0005] An artificial intelligence-based intelligent recommendation method for data services of a regional power marketing management information platform includes the following steps:

[0006] S1. Asynchronous data alignment: Collect user electricity usage behavior sequences and device alarm logs from the power marketing platform, and generate time-synchronized multi-dimensional data streams through event-driven timestamp resampling;

[0007] S2. Device Interference Stripping: Input the multi-dimensional data stream into a gated adversarial network, dynamically generate a mask matrix based on the device alarm level, strip away the distorted characteristics of power consumption caused by device anomalies, and output a user baseline demand vector.

[0008] S3. Service strategy mapping: performing similarity matching between the user baseline demand vector and the power service knowledge base to generate a set of recommended strategies.

[0009] Optionally, the S1 specifically includes:

[0010] S11, device alarm trigger collection: When the device alarm log is detected, the user's electricity usage behavior sequence within the set time window before and after the alarm trigger moment is synchronously captured to form a raw data segment;

[0011] S12, event stamp alignment: interpolate the voltage and current sampling points in the power usage behavior sequence to align their timestamps with the device alarm triggering time;

[0012] S13, interference intensity marking: according to the fault code in the alarm log, a predefined influence coefficient table is associated with each data segment to mark the device interference intensity level;

[0013] S14, multi-dimensional data reorganization: encapsulate the time-aligned power consumption behavior characteristics, equipment interference intensity level and user identity into a structured multi-dimensional data stream.

[0014] Optionally, the electricity consumption behavior characteristics are obtained based on the electricity consumption behavior sequence, and the electricity consumption behavior characteristics include the valley-peak electricity consumption ratio, the frequency of load mutation, and the steady-state duration.

[0015] Optionally, the S2 specifically includes:

[0016] S21, gating mask generation: inputting the multi-dimensional data stream into a gating generator, generating a dynamic binary mask matrix based on the device alarm level, wherein a high alarm level corresponds to a continuous weight decay channel of the mask matrix;

[0017] S22, feature correction operation: performing element-by-element multiplication of the mask matrix and the power usage behavior feature to suppress feature dimensions associated with device anomalies and generate a distortion suppression feature vector;

[0018] S23, adversarial verification: inputting the distortion suppression feature vector into the discriminator network, the discriminator simultaneously receives the reference feature vector under the normal state of the device, and calculates the feature consistency loss;

[0019] S24, demand vector reconstruction: The distortion suppression features are temporally modeled by stacking a bidirectional GRU network to output the user baseline demand vector.

[0020] Optionally, the forget gate weight of the bidirectional GRU network is negatively correlated with the device alarm level.

[0021] Optionally, the bidirectional GRU network specifically performs:

[0022] Receiving as input a distortion suppression feature sequence after interference suppression processing, wherein each time step includes a set of power consumption behavior feature data, constituting an input sequence of fixed time length;

[0023] The first layer of bidirectional gated recurrent unit network propagates the temporal state of the input sequence forward and backward respectively, extracting the context-related features of each time step in the forward and reverse directions;

[0024] The forward and reverse hidden states of the first layer’s bidirectional output are concatenated in the time step dimension to form an enhanced intermediate feature sequence, which serves as the input of the second layer network.

[0025] The second layer of bidirectional gated recurrent unit network performs bidirectional state propagation again on the intermediate feature sequence to model global time dependency information and enhance sequence modeling capabilities;

[0026] Perform a pooling operation on the forward and reverse hidden states of all time steps output by the second layer network, and calculate the average representation of its overall time dimension to aggregate the global features in the time series;

[0027] The pooled output representation is used as the user baseline demand vector and the final modeling result of this stage, which is used to describe the stable user electricity demand characteristics after interference stripping and timing modeling.

[0028] Optionally, the S3 specifically includes:

[0029] S31, knowledge base dynamic encoding: Parse each policy in the power service knowledge base into a service dimension vector, including a service timeliness label, an implementation cost coefficient, and an applicable scenario code. The applicable scenario code is dynamically weighted by the historical alarm frequency of the equipment.

[0030] S32, composite similarity calculation: performing multi-core matching on the user baseline demand vector and the service dimension vector, and calculating a weighted composite value of temporal similarity, cost sensitivity, and scenario compatibility;

[0031] S33, policy conflict resolution: When the composite similarity difference of multiple policies is less than the similarity threshold, the priority rearrangement based on the device health status is activated. When the current device failure rate is higher than the set value, the remote service policy is recommended first.

[0032] Optionally, the S3 also includes dynamic optimization of the queue: according to the real-time execution feedback data of the recommendation strategy, the weight coefficient in S32 is dynamically adjusted by using reinforcement learning to generate a set of recommendation strategies with valid timestamps.

[0033] Optionally, the calculation of the time series similarity includes: performing dynamic time warping on the electricity consumption behavior fluctuation sequence contained in the user baseline demand vector and the service supply cycle sequence defined in the service strategy, calculating the cumulative offset distance in the time series dimension through a sequence alignment algorithm, and mapping the cumulative offset distance to a normalized similarity value to measure the time series consistency between the user demand fluctuation pattern and the strategy service rhythm;

[0034] The cost sensitivity calculation includes: obtaining the user-side cost acceptance threshold and the implementation cost parameter of the service strategy, constructing a logarithmic growth function with the cost difference as the independent variable, measuring the difference in sensitive response between the two, and generating a normalized similarity index through the inverse transformation of the sensitive response difference value to reflect the degree of personalized matching of the strategy in the cost dimension;

[0035] The calculation of the scenario fit includes: counting the historical alarm frequencies of user devices in different categories to form a user device alarm frequency vector, and obtaining the applicable scenario coding vector marked by the corresponding service policy, evaluating the degree of linear correlation by calculating the Pearson correlation coefficient between the two, and converting the correlation value into a scenario fit index to measure the degree of consistency between the user's current device operating environment and the policy design adaptation scenario.

[0036] Optionally, the dynamic optimization of the queue specifically includes:

[0037] A reinforcement learning algorithm is built based on the execution feedback of recommendation strategies to achieve adaptive optimization of the recommendation strategy ranking mechanism. In the state representation, multi-dimensional indicators including user click behavior feedback, strategy execution time, and current device health status are introduced to form a state vector that characterizes the recommendation environment.

[0038] In the action selection phase, the similarity weight coefficient adjustment strategy is set based on the current state evaluation results. The adjustment range varies within a preset range to control the combined influence ratio of temporal similarity, cost sensitivity, and scenario fit, driving the real-time update of the strategy recommendation ranking.

[0039] After each action is executed, a reward function is constructed based on user response, strategy execution efficiency, and device status improvement effect to quantify the actual effect of the current recommended strategy combination and serve as a feedback signal for strategy update. Ultimately, through a continuous iterative reinforcement learning training process, the strategy convergence of the weight coefficient is completed, forming a recommendation queue optimization mechanism with dynamic perception capabilities.

[0040] Beneficial effects of the present invention:

[0041] The present invention solves the asynchronous sampling problem between electricity consumption behavior data and equipment status data through an event-driven data acquisition mechanism combined with a dynamic window selection and interpolation method triggered by equipment alarms. At the same time, a dynamic mask matrix is generated through a gated adversarial network to remove the disturbance of equipment anomalies on user electricity consumption behavior characteristics, and a bidirectional recurrent neural network is combined for time series modeling to construct a stable and robust user baseline demand vector, laying a data foundation for subsequent recommendation strategy matching, breaking through the technical bottleneck of severe model distortion in traditional systems under abnormal fluctuations.

[0042] The present invention introduces three types of heterogeneous indicators, namely temporal similarity, cost sensitivity and scenario fit, to construct a multi-core weighted matching system. During the matching process, the dynamic time warping algorithm, logarithmic cost function and Pearson correlation calculation are used respectively to comprehensively measure the matching degree between user needs and service strategies from three aspects: temporal behavior, economic burden and device status. This significantly improves the adaptability of recommendation results to individual differences and effectively overcomes the shortcomings of traditional label-based or static rule-based recommendation methods in personalization accuracy.

[0043] The present invention constructs a state space based on user feedback data and device health status, and uses recommendation click-through rate, strategy execution time and device improvement effect as reward factors. Through reinforcement learning, it dynamically updates the multi-core matching weight coefficient to form an intelligent recommendation queue that can self-adjust and continuously optimize. This mechanism enables the recommendation system to have real-time response capabilities to changes in the operating environment, thereby improving the accuracy and timeliness of service strategies in actual application. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 A flowchart of a recommended method according to an embodiment of the present invention is shown;

[0046] Figure 2 Schematic diagram of service strategy mapping according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0048] It should be noted that references in the specification to "one embodiment," "an embodiment," "an exemplary embodiment," "some embodiments," etc. indicate that the described embodiments may include specific features, structures, or characteristics, but not every embodiment necessarily includes such specific features, structures, or characteristics. In addition, when specific features, structures, or characteristics are described in conjunction with an embodiment, it is within the knowledge of persons skilled in the relevant art to implement such features, structures, or characteristics in conjunction with other embodiments (whether or not explicitly described).

[0049] In general, terms can be understood, at least in part, from their use in context. For example, depending at least in part on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending at least in part on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0050] like Figure 1-Figure 2 As shown, an artificial intelligence-based intelligent recommendation method for data services of a regional power marketing management information platform includes the following steps:

[0051] S1. Asynchronous data alignment: Collect user electricity usage behavior sequences and device alarm logs from the power marketing platform, and generate time-synchronized multi-dimensional data streams through event-driven timestamp resampling;

[0052] S2. Device Interference Stripping: Multi-dimensional data streams are fed into a gated adversarial network. A mask matrix is dynamically generated based on the device alarm level to strip away the distorted characteristics of power usage caused by device anomalies and output a user baseline demand vector.

[0053] S3, Service Strategy Mapping: Perform similarity matching between the user baseline demand vector and the power service knowledge base to generate a set of recommended strategies.

[0054] S1 specifically includes:

[0055] S11, device alarm trigger collection: When the device alarm log is detected, based on the alarm event timestamp t e , synchronously capture the time window set before and after the alarm is triggered [t e -T1,t e +T2], forming the original data fragment: Among them, t e Indicates the time when the device alarm is triggered. T1=30s indicates the warning time window, T2=5min indicates the delay time window, U i is the voltage value at the i-th moment, and multiple moments form a voltage sequence, I i is the current value at the i-th moment, and multiple moments form a current sequence.

[0056] S12, event stamp alignment: for the original data segments The time series of the sampling points in the Lagrange interpolation is performed to make it consistent with the alarm event timestamp t e For precise alignment, the interpolation function is expressed as:

[0057]

[0058] in, It represents the interpolation function, which estimates the value at the target time t. It is applicable to the interpolation of voltage series or current series. t represents the target interpolation time, which is the device alarm triggering time t e or a finely aligned time point nearby, t j is the jth known sampling moment, the sample point time used for interpolation, y j Indicates that j The corresponding observation value is the voltage U j or current I j , t m Divide t by the interpolation process j The other sampling time indexes except are used to construct the denominator term, and n is the total number of interpolation points, that is, the number of known data points used to construct the interpolation function (3 to 5).

[0059] S13, Interference intensity mark: Based on the fault code C recorded in the alarm log f , query the predefined influence coefficient table Φ and obtain the interference level L f and the influence coefficient γ f , and add the following markup to the current data fragment:

[0060] Among them, C f is the fault code (such as F301 means voltage sag), L f ∈{1,2,3,4,5} represents the interference intensity level (5 is the highest), γ f ∈[0.1,2.0] is the impact duration coefficient.

[0061] S14, multi-dimensional data reorganization: the electricity consumption behavior characteristics after time alignment F u , Equipment Interference Mark and user identity tag ID u Encapsulated as structured data stream: Among them, the electricity consumption behavior feature F u include:

[0062] Off-peak electricity consumption ratio R gp Defined as: in, For the defined low-peak period set (such as 0:006:00 at night), is the defined peak time set (such as 9:00-17:00 during the day), Δt is the sampling time interval, the numerator is the valley power consumption, and the denominator is the peak power consumption. The ratio of the two reflects the user's load preference and time characteristics. i is the active power, calculated as: P i =U i I i ;

[0063] Load mutation frequency F lc , represents the number of times the load changes exceed the set threshold per unit time, the load mutation frequency F lc Defined as: Among them, θ load is the mutation threshold (set according to industry experience, the default value is 0.5kW / s), is an indicative function. It takes 1 when the condition is met and 0 when it is not. It represents the number of times the load change rate exceeds the threshold in unit time, reflecting the volatility of the user's electricity consumption behavior. i represents the sampling point, and i=2 means starting from the second sampling point because steady-state judgment requires comparing two adjacent points.

[0064] Steady-state duration T ss , represents the total time duration of the continuous load change rate being lower than the set value, steady-state duration T ss Defined as: Among them, θ ss is the steady-state judgment threshold (the value is the same as the mutation threshold, 0.2kW / s). If the load change rate between two adjacent sampling points is lower than the steady-state judgment threshold, it is considered to be steady-state. The total time that the steady-state condition is met is counted to reflect the stability characteristics of the user's electricity consumption behavior.

[0065] Predefined influence coefficient table (example)

[0066]

[0067] S2 specifically includes:

[0068] S21, gated mask generation: multi-dimensional structured data stream S u =(ID u ,F u ,L f ,γ f ) Input gate generator, combined with the equipment alarm level L f Dynamically generate mask matrix M∈{0,1} d Or weighted mask matrix M′∈[0,1]d, for alarm level L f When ≥3, the continuous weight decay channel is activated, and the generation rule is as follows: M′=1-η·L f , where η∈[0.05,0.2] represents the weight attenuation coefficient, L f ∈{1,2,3,4,5} is the device interference level, the mask matrix dimension d is the same as the behavior feature vector F u One to one correspondence.

[0069] S22, feature correction operation: the mask matrix is combined with the original power consumption behavior feature F uPerform element-by-element multiplication to generate the distortion suppression feature vector Among them: ⊙ represents element-wise multiplication, is the eigenvector after distortion suppression. The suppressed channel is attenuated close to 0 at high interference levels and retains its original value at low levels.

[0070] S23, Adversarial Verification: Input the discriminator network D(·), and form a positive and negative sample pair with two comparison samples:

[0071] Positive sample pairs:

[0072] Negative sample pairs:

[0073] The discriminator loss function is a composite adversarial loss include:

[0074] Among them, KL(·||·) is the KL divergence of the feature distribution, which is used to maintain the distribution consistency of the corrected features. CE(·) is the cross entropy loss, which enhances the ability to identify abnormal samples. y∈{0,1}, where label 1 is normal and 0 is abnormal. α, β are loss weighting factors (set α=1.0, β=0.5).

[0075] S24, demand vector reconstruction: transform the distortion suppression feature sequence Input the stacked bidirectional GRU network BiGRU(·), perform time series modeling, and generate the user baseline demand vector D u , T is the time length, where the output dimension is compressed to 1 / 4 of the original feature vector;

[0076] The GRU forget gate is expressed as: f t =σ(W f ·[h t-1 ,x t ]+b f -η·L f ), where f t is the forget gate output, W f is the weight matrix, b f is the bias term, h t-1 is the hidden state at the previous moment, x t is the current input feature, L f is the alarm level of the current sample, η is the gate weight attenuation coefficient, and σ(·) is the sigmoid activation function. This design implements the gated inhibition mechanism of “the higher the interference level, the faster the forgetting”.

[0077] The specific scheme of the BiGRU (bidirectional GRU) network in reconstructing the user baseline demand vector is as follows:

[0078] 1. Input structure: The input sequence is a distortion suppression feature sequence

[0079] 2. Network structure

[0080] 1. Stacked bidirectional GRU structure: The network consists of two layers of bidirectional GRU units stacked together, expressed as:

[0081] in:

[0082] The first layer of bidirectional GRU:

[0083]

[0084]

[0085] The second layer of bidirectional GRU (the input is the concatenation of the forward and reverse hidden states of the first layer):

[0086]

[0087] Final output vector:

[0088] The forget gate of each GRU unit is expressed as: f t =σ(W f ·[h t-1 ,x t ]+b f -η·L f ).

[0089] 3. Output description:

[0090] The final output vector Du is used to represent the user's baseline power demand vector after removing device interference.

[0091] By default, the output dimension is set to 1 / 4 of the original feature vector to achieve compression and denoising of the original behavioral features.

[0092] S3 specifically includes:

[0093] S31, knowledge base dynamic encoding: parse each service strategy in the power service knowledge base into a service dimension vector: S k =[τ k ,c k ,s k ], where τ k is the service timeliness label, which indicates the average time required for the policy to be triggered and completed. k is the implementation cost coefficient, which reflects the average cost required for strategy execution (which can be money, resource consumption, etc.), kis the applicable scenario coding vector, which is dynamically weighted by the device's historical alarm frequency: Among them, λ i is the historical triggering frequency of the i-th type alarm, e i is the basic scenario vector corresponding to the i-th type of alarm, and n1 is the total number of alarm types.

[0094] S32, composite similarity calculation: calculate the user baseline demand vector D u and service strategy dimension vector S k The multi-core similarity of , including three types of sub-similarity and their weighted fusion:

[0095]

[0096] in:

[0097] Sim time To measure the time series similarity, the dynamic time warping (DTW) algorithm is used to align user behavior fluctuations with the policy service cycle. The user demand fluctuation sequence and the policy expected supply sequence are input and the DTW cumulative shortest path distance D is calculated. DTW , the similarity is calculated as: The smaller the DTW distance, the higher the similarity. This conversion form ensures that the similarity value is normalized in the range (0,1);

[0098] Sim cost is the cost sensitivity, which is equivalent to the logarithmic loss function log(1+|c u -c k |) is the inverse form of the price, which is used to measure the matching degree between the user's acceptance of the electricity price and the strategy cost, c u The cost sensitivity value of the user is c k is the strategy implementation cost, and the similarity is calculated as: When c k ≈c u When , the loss is minimum and the similarity is maximum; this form has a smooth descent characteristic, which suppresses the extreme impact of large cost differences on the evaluation results;

[0099] Sim scene Scenario fit, defined as the Pearson correlation coefficient between the alarm frequency of user devices and the coding of the scenario in which the policy is applicable;

[0100] enter:

[0101] User equipment alarm frequency vector λ=[λ1,λ2,...,λ m ];

[0102] Strategy applicable scenario encoding vector s k =[s1,s2,...,s m ];

[0103] Similarity calculation: in, is the average frequency of user alarms, The encoding mean of the strategy scenario is in the range of [-1, 1], which can be converted to the interval of [0, 1] for similarity fusion through linear normalization.

[0104] δ, γ is the multi-core weight coefficient, which takes values of 0.5, 0.3, and 0.2 respectively, satisfying

[0105] S33, strategy conflict resolution: When the difference in the composite similarity values of multiple strategies is less than the set similarity threshold π ( π takes a value of 0.05, that is, within 5%. When the composite similarity difference between two strategies is less than 0.05, they are considered to have equal priority, triggering the conflict resolution mechanism:

[0106] Calculate the current device health index H: Among them, f b is the monthly triggering frequency of the b-type alarm, w b represents the severity weight of the b-th type alarm, H∈[0,1] represents the equipment health index, and the smaller the value, the less healthy the equipment. When H<0.7, the priority of remote service strategies (such as online repair) is increased by 3 levels to avoid on-site processing risks.

[0107] S34, dynamic optimization of queues: For the recommended strategy set, based on the real-time execution feedback of the strategy, a reinforcement learning optimization algorithm is constructed to dynamically adjust the similarity weight coefficient of each core.

[0108] The state space of the reinforcement learning optimization algorithm is represented as: s t =[CTR t ,T t ,H t ], where CTR t is the current click rate, T t is the execution time of the current strategy, H t is the current device health index;

[0109] The action space of the reinforcement learning optimization algorithm is: adjust the three similarity weights δ, γ, the adjustable range is ±15% each time;

[0110] Reward function: Among them, T is the actual strategy time, T max is the maximum allowed policy response time, H is the current device health index, and R is the comprehensive incentive value, which is used to feedback the policy recommendation effect.

[0111] In the above content, the user baseline demand vector is essentially a set of structured features extracted from the user's historical electricity consumption behavior, such as load fluctuation, off-peak preference, steady-state level, etc., which represents the user's energy demand preference.

[0112] Service dimension vector: represents the key attributes of each service policy during execution, such as response time, cost impact, and applicable device failure context. It is actually an abstraction of the service policy.

[0113] Both are essentially vector forms of multi-dimensional attributes and have comparable semantic structures, making similarity matching symmetrical and meaningful in the expression space.

[0114] The introduction of multi-core similarity (temporal pattern, cost sensitivity, and scenario fit) solves the problem that a single distance calculation cannot cover heterogeneous features in multiple dimensions;

[0115] User behavior data comes from smart meters, load monitoring terminals, energy efficiency analysis systems, etc., and has high frequency and continuity;

[0116] Equipment fault data and alarm logs have been widely deployed in the power information platform, supporting the construction of equipment health models and scenario mapping;

[0117] The service strategy library is maintained by the power enterprise platform and includes fields such as response rules, processing methods, resource consumption, etc., and has structured expression conditions.

[0118] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0119] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based intelligent recommendation method for data services of a regional power marketing management information platform, characterized in that: The following steps are involved: S1. Asynchronous data alignment: Collect user electricity usage behavior sequences and device alarm logs from the power marketing platform, and generate time-synchronized multi-dimensional data streams through event-driven timestamp resampling; S2. Device Interference Stripping: Input the multi-dimensional data stream into a gated adversarial network, dynamically generate a mask matrix based on the device alarm level, strip away the distorted characteristics of power consumption caused by device anomalies, and output a user baseline demand vector. S3. Service strategy mapping: performing similarity matching between the user baseline demand vector and the power service knowledge base to generate a set of recommended strategies.

2. The method for intelligent recommendation of data services for regional power marketing management information platform based on artificial intelligence according to claim 1, characterized in that: Said S1 specifically includes: S11, device alarm trigger collection: When the device alarm log is detected, the user's electricity usage behavior sequence within the set time window before and after the alarm trigger moment is synchronously captured to form a raw data segment; S12, event stamp alignment: interpolate the voltage and current sampling points in the power usage behavior sequence to align their timestamps with the device alarm triggering time; S13, interference intensity marking: according to the fault code in the alarm log, a predefined influence coefficient table is associated with each data segment to mark the device interference intensity level; S14, multi-dimensional data reorganization: encapsulate the time-aligned power consumption behavior characteristics, equipment interference intensity level and user identity into a structured multi-dimensional data stream.

3. The method for intelligently recommending data services for a regional power marketing management information platform based on artificial intelligence according to claim 2, characterized in that: The electricity consumption behavior characteristics are obtained based on the electricity consumption behavior sequence, and the electricity consumption behavior characteristics include the off-peak electricity consumption ratio, the frequency of load mutation, and the steady-state duration.

4. The method for intelligent recommendation of data services for a regional power marketing management information platform based on artificial intelligence according to claim 1, characterized in that: The S2 specifically includes: S21, gating mask generation: inputting the multi-dimensional data stream into a gating generator, generating a dynamic binary mask matrix based on the device alarm level, wherein a high alarm level corresponds to a continuous weight decay channel of the mask matrix; S22, feature correction operation: performing element-by-element multiplication of the mask matrix and the power usage behavior feature to suppress feature dimensions associated with device anomalies and generate a distortion suppression feature vector; S23, adversarial verification: inputting the distortion suppression feature vector into the discriminator network, the discriminator simultaneously receives the reference feature vector under the normal state of the device, and calculates the feature consistency loss; S24, demand vector reconstruction: The distortion suppression features are temporally modeled by stacking a bidirectional GRU network to output the user baseline demand vector.

5. The method for intelligent recommendation of data services for regional power marketing management information platform based on artificial intelligence according to claim 4 is characterized in that: The forget gate weight of the bidirectional GRU network is negatively correlated with the device alarm level.

6. The method for intelligent recommendation of data services for regional power marketing management information platform based on artificial intelligence according to claim 5, characterized in that: The bidirectional GRU network specifically performs: Receiving as input a distortion suppression feature sequence after interference suppression processing, wherein each time step includes a set of power consumption behavior feature data, constituting an input sequence of fixed time length; The first layer of bidirectional gated recurrent unit network propagates the temporal state of the input sequence forward and backward respectively, extracting the context-related features of each time step in the forward and reverse directions; The forward and reverse hidden states of the first layer’s bidirectional output are concatenated in the time step dimension to form an enhanced intermediate feature sequence, which serves as the input of the second layer network. The second layer of bidirectional gated recurrent unit network performs bidirectional state propagation again on the intermediate feature sequence to model global time dependency information and enhance sequence modeling capabilities; Perform a pooling operation on the forward and reverse hidden states of all time steps output by the second layer network, and calculate the average representation of its overall time dimension to aggregate the global features in the time series; The pooled output representation is used as the user baseline demand vector and the final modeling result of this stage, which is used to describe the stable user electricity demand characteristics after interference stripping and timing modeling.

7. The method for intelligently recommending data services for a regional power marketing management information platform based on artificial intelligence according to claim 1, characterized in that: The S3 specifically includes: S31, knowledge base dynamic encoding: Parse each policy in the power service knowledge base into a service dimension vector, including a service timeliness label, an implementation cost coefficient, and an applicable scenario code. The applicable scenario code is dynamically weighted by the historical alarm frequency of the equipment. S32, composite similarity calculation: performing multi-core matching on the user baseline demand vector and the service dimension vector, and calculating a weighted composite value of temporal similarity, cost sensitivity, and scenario compatibility; S33, policy conflict resolution: When the composite similarity difference of multiple policies is less than the similarity threshold, the priority rearrangement based on the device health status is activated. When the current device failure rate is higher than the set value, the remote service policy is recommended first.

8. The method for intelligently recommending data services for a regional power marketing management information platform based on artificial intelligence according to claim 7, characterized in that: The S3 also includes dynamic optimization of the queue: according to the real-time execution feedback data of the recommendation strategy, the weight coefficient in S32 is dynamically adjusted by using reinforcement learning to generate a set of recommendation strategies with valid timestamps.

9. The method for intelligently recommending data services for a regional power marketing management information platform based on artificial intelligence according to claim 7, characterized in that: The calculation of the temporal similarity includes: performing dynamic time warping on the electricity consumption behavior fluctuation sequence contained in the user baseline demand vector and the service supply cycle sequence defined in the service strategy, calculating the cumulative offset distance in the temporal dimension through a sequence alignment algorithm, and mapping the cumulative offset distance to a normalized similarity value to measure the temporal consistency between the user demand fluctuation pattern and the strategy service rhythm; The cost sensitivity calculation includes: obtaining the user-side cost acceptance threshold and the implementation cost parameter of the service strategy, constructing a logarithmic growth function with the cost difference as the independent variable, measuring the difference in sensitive response between the two, and generating a normalized similarity index through the inverse transformation of the sensitive response difference value to reflect the degree of personalized matching of the strategy in the cost dimension; The calculation of the scenario fit includes: counting the historical alarm frequencies of user devices in different categories to form a user device alarm frequency vector, and obtaining the applicable scenario coding vector marked by the corresponding service policy, evaluating the degree of linear correlation by calculating the Pearson correlation coefficient between the two, and converting the correlation value into a scenario fit index to measure the degree of consistency between the user's current device operating environment and the policy design adaptation scenario.

10. The method for intelligently recommending data services for a regional power marketing management information platform based on artificial intelligence according to claim 8, characterized in that: The dynamic optimization of the queue specifically includes: A reinforcement learning algorithm is built based on the execution feedback of recommendation strategies to achieve adaptive optimization of the recommendation strategy ranking mechanism. In the state representation, multi-dimensional indicators including user click behavior feedback, strategy execution time, and current device health status are introduced to form a state vector that characterizes the recommendation environment. In the action selection phase, the similarity weight coefficient adjustment strategy is set based on the current state evaluation results. The adjustment range varies within a preset range to control the combined influence ratio of temporal similarity, cost sensitivity, and scenario fit, driving the real-time update of the strategy recommendation ranking. After each action is executed, a reward function is constructed based on user response, strategy execution efficiency, and device status improvement effect to quantify the actual effect of the current recommended strategy combination and serve as a feedback signal for strategy update. Ultimately, through a continuous iterative reinforcement learning training process, the strategy convergence of the weight coefficient is completed, forming a recommendation queue optimization mechanism with dynamic perception capabilities.

Images