CN-121052443-B - Hierarchical system energy consumption prediction method and system based on machine learning

Abstract

The invention discloses a grading system energy consumption prediction method and system based on machine learning, in particular to the field of processing energy consumption prediction, which are used for solving the problem of grading energy consumption prediction deviation of a powder processing production line, wherein a multi-source time sequence account book is constructed to pull a dried and graded vein into a traceable line, disturbance association and key events clearly exchange the arrival and the arrival of energy consumption fluctuation, and joint prediction is not stopped in static experience any more, and real load is depicted by acting with a screen surface along a propagation path; the energy consumption impact degree is comprehensively obtained through time difference density and accumulation release work amount, the cost is generated on the operation side to influence the label and time slot priority, the order energy consumption is layered to be close to the site rhythm, prediction and accounting are more consistent, on-line checking is carried out, small-step revising is carried out on alignment deviation, short-time fluctuation convergence is carried out, long-period cost curves tend to be smooth, clear evidence chains are reserved, check is carried out, stable response is still kept under the alternation of batches and beats, and settlement deviation caused by misjudgment is reduced.

Inventors

• WANG LIANG

Assignees

• 上海晨洋新材料有限公司

Dates

Publication Date

20260508

Application Date

20250825

Claims (6)

1. 1. The hierarchical system energy consumption prediction method based on machine learning is characterized by comprising the following steps: s1, aligning a time axis with batch labeling, and collecting material properties, control instructions and metering electric energy of each processing link to generate a multi-source time sequence account book; s2, establishing disturbance association with the position according to the sequence, extracting a conducting path of moisture fluctuation, granularity migration and beat change, forming a key event sequence pointing to the classified energy consumption, and marking arrival time and source; s3, constructing an order-level feature sequence by using a key event sequence, inputting a time sequence learning model, outputting a combined prediction of classified short-time energy consumption and unit output energy consumption, and generating an interpretation result affecting contribution; S4, before the output of the operation side is generated, the penetration time difference density and the accumulation release work load are calculated for the key event fragments of the current batch, the energy consumption impact degree is obtained through the sequence ratio learning and discrimination, the combined prediction and interpretation result and the energy consumption impact degree are mapped to an electricity price curve and the productivity constraint, and the cost influence label, the time slot priority and the order energy consumption layering are generated; S5, monitoring deviation of actual energy consumption and joint prediction, carrying out small-step iteration revising on a time sequence learning model and disturbance association according to a deviation direction, updating labels and layering results in a multi-source time sequence account book, and taking effect in the next batch; S4, selecting a current batch of key event fragments from a key event sequence, calculating penetration time difference density and stacking release work load, comparing and distinguishing the penetration time difference density and the stacking release work load through order ratio, inputting the penetration time difference density and the stacking release work load, comparing relative orders, outputting energy consumption impact degree, mapping a combined prediction, a classified short-time energy consumption prediction value, a unit output energy consumption prediction value, an interpretation result and the energy consumption impact degree to an electricity price curve and productivity constraint, generating a cost influence label through multiplying an electricity price factor, generating a time slot priority according to productivity constraint sequencing time slots, and generating order energy layering based on an energy consumption impact degree layering order; The penetration time difference density obtaining logic is used for establishing a time difference sequence by marking a flow track rising edge event of a feeding beat batch sequence and a downstream fine particle concentration track transition event of a grading amplitude batch sequence, taking the reciprocal to generate a transmission rate sequence, counting the rate event intensity exceeding a physical lower limit threshold, and dividing the rate event intensity by the observation window duration; The pile-up release work amount obtaining logic is used for integrating and accumulating excess power through positioning a release interval sequence of vibration acceleration of a grading amplitude batch sequence and active power phase shift of a metering electric energy batch sequence, dividing the excess power by the mass flow of a feeding beat batch sequence and summing the excess power; Step S5 comprises the steps of extracting actual energy consumption values of a current batch grading unit from a metered electric energy batch sequence, comparing the actual energy consumption values with grading short-time energy consumption predicted values and unit output energy consumption predicted values, calculating deviation, determining deviation directions, carrying out small-step iteration adjustment on conversion matrixes and bias parameters of a long-short-period memory network according to the deviation directions, revising conduction path tolerance ranges of moisture fluctuation events, granularity migration events and beat change events in a disturbance association adjustment key event sequence, updating cost influence labels, time slot priorities and order energy consumption layering results in a multi-source time sequence account book, recalculating labels based on the revised predicted values multiplied by electric price factors, reordering time slots and layering orders according to updated energy consumption impact degrees, and storing revising results to take effect in the next batch.
2. 2. The machine learning based hierarchical system energy consumption prediction method of claim 1, wherein: Step S1 comprises the steps of collecting dry material moisture content records, feeding flow beat records, screening particle size distribution records, classified vibration amplitude records and air volume records, conveying linear speed records, control command signals and metering electric energy readings, aligning time axes, converting all time stamp sequences by taking a unified clock of a production line master control system as a reference, marking batches, marking all batch sequences by taking a flow rising edge starting point as a demarcation mark, integrating collected data into a unified data set according to the sequence of the reference time axes, and generating a multi-source time sequence account book.
3. 3. The machine learning based hierarchical system energy consumption prediction method of claim 2, wherein: Step S2 comprises the steps of extracting each batch sequence from a multi-source time sequence account book, arranging the batch sequences according to a reference time axis, establishing disturbance association, identifying moisture fluctuation events, granularity migration events and beat change events according to the production line link sequence and physical positions, linking the moisture fluctuation events, the granularity migration events and the beat change events to form a conducting path, extracting a conducting path tracking response point, calculating propagation delay, integrating path segments lower than a physical lower bound to form a conducting path set, forming a key event sequence, screening disturbance events pointing to a power peak value of a grading unit in the metering electric energy batch sequence from the conducting path set, marking the arrival time and the grading arrival time of each event in the source record key event sequence on the reference time axis, and marking the source link name.
4. 4. A machine learning based hierarchical system energy consumption prediction method according to claim 3, characterized in that: where the physical lower bound represents the shortest theoretical time limit for the disturbance signal to propagate from the upstream ring node to the downstream.
5. 5. A machine learning based hierarchical system energy consumption prediction method according to claim 3, characterized in that: Step S3 comprises the steps of extracting an event value sequence, a time sequence and a source sequence from a key event sequence, constructing a order level feature sequence, calculating a moisture accumulation feature, a granularity accumulation feature, a beat accumulation feature and interval difference value array according to a batch number group, inputting a long-short-period memory network model processing sequence, outputting a classified short-period energy consumption predicted value and a unit output energy consumption predicted value to form a combined prediction, generating an interpretation result of influence contribution, calculating the contribution score of each order level feature sequence element by using an attention mechanism, and listing the influence contribution of each source to the combined prediction by associating the source sequence.
6. 6. A machine learning based hierarchical system energy consumption prediction system for implementing the machine learning based hierarchical system energy consumption prediction method of any one of claims 1-5, comprising: The data integration module is responsible for aligning a time axis with batch labeling, collecting material properties, control instructions and metering electric energy of each processing link, and generating a multi-source time sequence account book so as to establish a unified time sequence data base; The disturbance path extraction module is used for establishing disturbance association with the position according to the sequence, extracting a conducting path with moisture fluctuation, granularity migration and beat change, forming a key event sequence pointing to the classified energy consumption, and marking arrival time and source so as to capture disturbance propagation dynamics; The prediction and interpretation module builds an order-level feature sequence by using a key event sequence, inputs a time sequence learning model, outputs the combined prediction of the classified short-time energy consumption and the unit output energy consumption, and generates an interpretation result of influence contribution so as to realize dynamic prediction and cause analysis of the energy consumption; Before generating operation side output, the operation output generation module calculates the penetration time difference density and the accumulation release work load according to the key event fragments of the current batch, obtains the energy consumption impact degree through sequence comparison learning and discrimination, maps the combined prediction and interpretation result and the energy consumption impact degree to an electricity price curve and productivity constraint, generates a cost influence label, a time slot priority and order energy consumption layering, and predicts and makes a business decision through a bridging technology; The iterative optimization module monitors the deviation of the actual energy consumption and the joint prediction, revises the time sequence learning model and the disturbance association in a small step iteration mode according to the deviation direction, updates the labels and layering results in the multi-source time sequence account book, and takes effect in the next batch so as to ensure the continuous adaptability and accuracy of the model.

Description

Hierarchical system energy consumption prediction method and system based on machine learning Technical Field The invention relates to the field of processing energy consumption prediction, in particular to a hierarchical system energy consumption prediction method and system based on machine learning. Background In the aluminum hydroxide powder processing production line, raw materials sequentially enter a classification unit through links of drying, feeding, screening, conveying and the like. The moisture content and the granularity of the materials are different from batch to batch, and the regulation strategy and the beat of an upstream link are not constant, so that the flow, the material layer thickness and the screen surface load can be changed. The grading unit usually works cooperatively by vibration and air quantity, the power consumption of the motor and the fan synchronously fluctuates along with the feeding state and the control command, and hysteresis effects are generated in the feeding return loop and the stacking-evacuating process. As a result, the classification unit is in an environment continuously excited by upstream disturbances, the power curve presents a trace of short-time spikes and medium-time rollbacks, and the energy consumption level changes rapidly with the details of the working conditions. However, the current energy consumption prediction is based on static experience or a fixed statistical model of single equipment, the coupling propagation between 'dry output → feeding beat → screening transmission → return loop → air quantity and amplitude command' is not characterized, the phase dislocation caused by control lag and self-organizing accumulation of materials is ignored, and the disturbance from the source to the classification unit is amplified and overlapped in a dislocation manner, so that short-time prediction frequently deviates from the real load. The deviation not only affects the adjusting time of the air quantity and the amplitude, but also easily triggers unnecessary gear jump and start-stop, further changes the returning proportion and the granularity distribution, and forms a repeated amplified closed loop error. In order to solve the above problems, a technical solution is now provided. Disclosure of Invention In order to overcome the defects in the prior art, the embodiment of the invention provides a grading system energy consumption prediction method and system based on machine learning, which are characterized in that a multi-source time sequence account book is constructed to draw a line which can be traced from a dry to graded context, disturbance association and key events clearly cross the incoming and outgoing routes of energy consumption fluctuation, joint prediction is not stopped in static experience, and the real load is marked by working with a screening surface along a propagation path, the energy consumption impact degree is comprehensively obtained by penetrating time difference density and stacking release work quantity, the management side generates a cost influence label and a time slot priority according to the energy consumption impact degree, the order energy consumption is layered close to a site rhythm, the prediction and the accounting are more consistent, on-line check is carried out to revise alignment deviation in small steps, short-time fluctuation is converged, long-period cost curves tend to be smooth, clear evidence chains are reserved at the same time, stable response is still maintained under the alternation of batches and beats, settlement deviation and management friction caused by misjudgment are reduced, and the problems in the background technology are solved. In order to achieve the above purpose, the present invention provides the following technical solutions: the hierarchical system energy consumption prediction method based on machine learning comprises the following steps: s1, aligning a time axis with batch labeling, and collecting material properties, control instructions and metering electric energy of each processing link to generate a multi-source time sequence account book; s2, establishing disturbance association with the position according to the sequence, extracting a conducting path of moisture fluctuation, granularity migration and beat change, forming a key event sequence pointing to the classified energy consumption, and marking arrival time and source; s3, constructing an order-level feature sequence by using a key event sequence, inputting a time sequence learning model, outputting a combined prediction of classified short-time energy consumption and unit output energy consumption, and generating an interpretation result affecting contribution; S4, before the output of the operation side is generated, the penetration time difference density and the accumulation release work load are calculated for the key event fragments of the current batch, the energy consumption impact degree is obtained through