CN111401651A - A method for predicting material procurement demand based on LSTM network - Google Patents

Abstract

The invention discloses a method for predicting material purchasing demand based on a L STM network, which comprises the following steps of S1, processing data acquired in advance by adopting a preset method, S2, comparing an original prediction model with a recurrent neural network model by adopting the preset method, S3, constructing a L STM model by adopting the preset method, taking the processed data as the input of a L STM model, and obtaining a predicted value.

Description

A method for predicting material procurement demand based on LSTM network

technical field

The invention relates to the technical field of methods for predicting material procurement requirements, in particular to a method for predicting material procurement requirements based on an LSTM network.

Background technique

At present, with the rapid development of my country's society and economy, various functional departments, such as the power department, have higher and higher demand for materials, which has promoted the prosperity of the power grid engineering market, and has also brought greater challenges to related enterprises. Only by further optimizing enterprise management And various resource allocation, improve resource utilization and engineering design and development efficiency, in order to adapt to the new market situation and meet various challenges.

Taking the power sector and enterprises as an example, how to accurately predict the material needs of substations and distribution network projects, improve capital utilization and save costs on the premise of ensuring the progress of the project, is of great significance and helps to improve the competitiveness of these departments and enterprises. .

In the aspect of material demand forecasting, the existing technology generally uses the procurement material demand forecasting model for forecasting. The model first obtains the actual planned procurement of the project by sorting out the feasibility study approval data and comprehensive planning data in the past three years, as well as the historical purchase orders for project materials. Attribute information such as amount, actual purchase amount, voltage level, and related sub-category purchase status, and then cluster the projects through the project attribute information, and confirm the project type to which the clustering result belongs, and then take the total investment amount and project type as the starting point At the same time, taking into account factors such as annual growth rate, project approval factors, material unit price fluctuations, equipment use status and other factors, a regression analysis was carried out on the purchase volume of small categories, and insignificant factors were eliminated to obtain multi-factor regression coefficients. Finally, through the obtained The multiple regression model and the values of related factors in the future predict the demand for various materials procurement.

However, the existing procurement material demand prediction model has the following defects: 1. The data is redundant and discrete, and the purchase volume of items fluctuates greatly. The procurement data is not cleaned, and the impact of outliers on the prediction is ignored; 2. The procurement There are too many influencing factors of demand, and the ordinary linear regression model cannot fit the prediction function well; 3. The time series relationship between historical purchases is not considered, and the prediction accuracy is low.

For the problems in the related technologies, no effective solutions have been proposed so far.

SUMMARY OF THE INVENTION

(1) Technical problems solved

In view of the deficiencies of the prior art, the present invention provides a method for predicting the demand for material procurement based on the LSTM network, which has high prediction accuracy, and can achieve the purposes of procurement optimization and cost optimization, thereby solving the problems in the background art.

(2) Technical solutions

In order to achieve the above-mentioned purposes of having high prediction accuracy and realizing procurement optimization and cost optimization, the specific technical solutions adopted in the present invention are as follows:

A method for predicting material procurement requirements based on LSTM network, comprising the following steps:

S1: use a preset method to perform data processing on the pre-collected data;

S2: Compare the original prediction model and the recurrent neural network model by a preset method;

S3: Construct an LSTM model by using a preset method, and use the processed data as the input of the LSTM model to obtain a predicted value.

Preferably, the step S1 adopts a preset method to perform data processing on the pre-collected data, which specifically includes the following steps:

S11: Collect the required data through the prepared database to obtain the collected data;

S12: Perform data cleaning on the collected data by a preset method to obtain processed data. Through the above processing, a more objective and real purchase amount can be obtained.

Preferably, the step S12 performs data cleaning on the collected data by a preset method, and obtaining the processed data specifically includes the following steps:

S121: Perform random sampling and estimation on the original time series data in the collected data to obtain a plurality of pieces of estimation data, and fill up the gaps and defects generated by the random sampling to obtain a plurality of pieces of fill-in estimation data;

S122: Classify all the supplemented estimated data according to the sampling time point, obtain multiple groups of time-classified data, and sort each group of the time-classified data according to size to obtain multiple groups of sorted arrays;

S123: Process multiple groups of the sorted arrays to obtain a plurality of corresponding average data, and form an average sequence by using the plurality of average data;

S124: Outputting the mean value sequence, that is, completing the removal of the abnormal data of the purchase amount, and obtaining the processed data. Through the above processing, the abnormally high and low purchase volume is effectively removed, so that the data has low discreteness and certain regularity.

Preferably, the step S2 compares the original prediction model and the cyclic neural network model by a preset method, and specifically includes the following steps: comparing the original prediction model and the cyclic neural network model from four data processing, functional relationship, time dependence, and trend. A comparative analysis is carried out in terms of aspects, and a more suitable prediction model is found according to the defects of the original prediction model. Through the above processing, a more suitable prediction model can be found for the defects of the original prediction model.

Preferably, the step S3 adopts a preset method to construct an LSTM model, and uses the processed data as the input of the LSTM model, and obtaining the predicted value specifically includes the following steps:

S31: establish an LSTM model by a preset method;

S32: Use the processed data as the input sequence of the LSTM model, and mark it as {x ₁ , x ₂ ,...x _t ...}, where x _t represents the input at time t;

S33: Train the _LSTM model to obtain forget gate weight W _f , forget gate bias b _f , input gate weight Wi , input gate bias _bi , cell weight W _c , cell bias b _c , and output gate weight W _o and output gate bias b _o ;

S34: After the LSTM model is trained, input the processed data at time t into the trained LSTM model, and obtain the data at time t+1, which is the predicted value. Through the above processing, an LSTM model can be obtained, and a predicted value can be obtained through the LSTM model.

Preferably, the step S31 to establish an LSTM model by a preset method specifically includes the following steps:

S311: Selectively forget the input information from the previous node, and obtain the forgetting threshold f _t+1 as the forgetting gate by calculating, to control which c _t-1 of the previous state need to be forgotten, and the forgetting threshold f _t+1 The expression is:

f _t+1 =σ(W _f ·[h _t , x _t ]+b _f );

Among them, h _t represents the output of the unit at time t, W _f represents the weight of the forget gate, b _f represents the bias of the forget gate, x _t represents the input at time t, and σ represents the sigmod function;

S312: Selectively memorize the input information, use i _t+1 to represent the input threshold, and use the cell state C _t at time t to control the control signal of the selection gate, input the threshold it _{+ 1} and the cell state at time t The expressions of C _t are:

i _t+1 =σ(W _i ·[h _t , x _t ]+b _i ), C _t =tanh(W _c ·[h _t , x _t ]+b _c );

Among them, Wi represents the weight of the input gate, _bi represents the bias of the input gate, W _c represents the weight of the cell, _{and bc represents} the bias of the cell;

S313: Update the state C _t of the cell at time t to obtain the state C _t+1 of the cell at time t+1, and the expression of the state C _t+1 of the cell at time t+1 is:

C _t+1 =f _t+1 *C _t +i _t+1 *C _t ;

S314: Controlled by the output threshold O _t+1 , and scaling the O _t obtained in the previous stage to obtain the output h _{t+1 at time t+1} , the output thresholds O _t+1 and the output at time t+1 The expressions of h _t+1 are:

O _t+1 =σ(W _o ·[h _t , x _t ]+b _o ), h _t+1 =O _t+1 *tanh(C _t+1 );

where W _o represents the output gate weight, and b _o represents the output gate bias. Through the above processing, the construction of the LSTM model can be realized.

Preferably, the training of the LSTM model in the step S33 further includes the following steps: when training the LSTM model, the training is performed in cycles of 1 day, 7 days, 8 days, 10 days and 12 days respectively, and then The model with the best training effect is selected for prediction. Through the above processing, better training results can be obtained.

Preferably, the training of the LSTM model in the step S33 further includes the following steps:

In the case of insufficient training, the training effect can be achieved by increasing the nodes in the network or increasing the training period of the network;

In the case of overfitting, the training period can be reduced or controlled, and the training of the network is stopped before the inflection point of the data to achieve the training effect. Through the above processing, the training effect of the LSTM model is guaranteed.

(3) Beneficial effects

Compared with the prior art, the present invention provides a method for predicting the demand for material procurement based on the LSTM network, which has the following beneficial effects:

(1) The present invention processes the data through data cleaning, and removes the abnormally high and low purchase volume, so that the data has low discreteness and certain regularity, thereby effectively avoiding the impact of abnormal values on prediction, improving its prediction accuracy.

(2) The present invention constructs the LSTM model by comparing the original prediction model and the cyclic neural network model, and can better fit the complex joint distribution through the nonlinear relationship, and map the relationship between the input and output well, so that the prediction model can It is flexibly applied to the responsible joint distribution to ensure its prediction effect.

(3) In the present invention, by constructing and training a material procurement prediction model based on LSTM network, considering the time series relationship between historical procurement quantities, and increasing or decreasing the level according to the time series, the model can be repeated over time, improving the The prediction accuracy of the model.

(4) The present invention uses the updated cell state at time t-1 and the output of the cell as well as the data at time t as input. Through the combination of three inputs and one output, the LSTM model can remember the long-term state. Therefore, it can effectively solve the long-distance dependence problem.

(5) The present invention verifies and guides the collection of traditional procurement requirements through the procurement demand prediction model, and guides to the final step-by-step replacement and IT solidification, so that the procurement demand management work of telecom operators is scientific, refined and efficient, and the procurement process releases manpower, Improve efficiency, accurately predict procurement needs, lean management, and carry out procurement in a planned way, effectively reducing sudden procurement costs, warehousing costs and engineering time costs, and achieve procurement optimization and cost optimization.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings required in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some of the present invention. In the embodiments, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without any creative effort.

Fig. 1 is a flowchart of a method for predicting material procurement demand based on an LSTM network according to an embodiment of the present invention;

2 is a data visualization diagram before and after cleaning of a 10kv transformer in a method for predicting material procurement requirements based on an LSTM network according to an embodiment of the present invention;

3 is a data visualization diagram before and after cleaning of 10kv cable terminal procurement data in the method for predicting material procurement requirements based on an LSTM network according to an embodiment of the present invention;

4 is a schematic structural diagram of an LSTM model structure in a method for predicting material procurement requirements based on an LSTM network according to an embodiment of the present invention;

5 is a schematic diagram of a training result of a 10kv transformer in a method for predicting material procurement requirements based on an LSTM network according to an embodiment of the present invention;

6 is a schematic diagram of a prediction result of a 10kv cable terminal in a method for predicting material procurement requirements based on an LSTM network according to an embodiment of the present invention;

7 is a schematic diagram of a prediction result of a low-voltage switchgear in a method for predicting material procurement requirements based on an LSTM network according to an embodiment of the present invention;

8 is a schematic diagram of a prediction result of a power cable in a method for predicting material procurement requirements based on an LSTM network according to an embodiment of the present invention;

9 is a schematic diagram of a prediction result of an overhead insulated wire in a method for predicting material procurement requirements based on an LSTM network according to an embodiment of the present invention;

FIG. 10 is a schematic diagram of the prediction result of the grounding line in the method for predicting the demand for material procurement based on the LSTM network according to an embodiment of the present invention.

Detailed ways

In order to further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention, and are mainly used to illustrate the embodiments, and can be used in conjunction with the relevant descriptions in the specification to explain the operation principles of the embodiments. For these, those of ordinary skill in the art will understand other possible implementations and the advantages of the present invention. Components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.

According to an embodiment of the present invention, a method for predicting the demand for material procurement based on an LSTM network is provided.

The present invention will now be further described with reference to the accompanying drawings and specific embodiments. As shown in Figures 1-10, the method for predicting the demand for material procurement based on an LSTM network according to an embodiment of the present invention includes the following steps:

S1: use a preset method to perform data processing on the pre-collected data;

Wherein, the S1 specifically includes the following steps:

S11: Collect the required data through the prepared database to obtain the collected data;

Specifically, the items selected in this embodiment are items with large purchase volume and frequent purchase cycles, including 6 types, namely 10kv transformers, 10kv cable terminals, low-voltage switch cabinets, power cables, overhead insulated conductors, and grounding lines. Among them, 10kv transformer and 10kv cable terminal are selected for data visualization display. Because the data fluctuates greatly and the discreteness is too high, it is not friendly enough for the establishment of the procurement prediction model. Therefore, data cleaning is required, and the purchase volume is abnormally high and low. Data deletion, as shown in Figure 2 is the data visualization diagram before and after cleaning the 10kv transformer, and Figure 3 is the data visualization diagram before and after cleaning the 10kv cable terminal procurement data.

The data in this example comes from the 2014-2019 harvest data of the State Grid Materials Department. According to an article of the State Grid Shanghai Electric Power Company, the proportion of the project clustering overall distribution table of the research on agreement inventory demand forecasting based on big data analysis is selected. Top 6 species, and collate the dates of the 6 species and their corresponding purchases from the 2014-2019 harvest data table.

S12: Perform data cleaning on the collected data by a preset method to obtain processed data.

Specifically, the S12 specifically includes the following steps:

S121: Perform random sampling and estimation on the original time series data in the collected data to obtain a plurality of pieces of estimation data, and fill up the gaps and defects generated by the random sampling to obtain a plurality of pieces of fill-in estimation data;

S122: Classify all the supplemented estimated data according to the sampling time point, obtain multiple groups of time-classified data, and sort each group of the time-classified data according to size to obtain multiple groups of sorted arrays;

S123: Process multiple groups of the sorted arrays to obtain a plurality of corresponding average data, and form an average sequence by using the plurality of average data;

S124: Outputting the mean value sequence, that is, completing the removal of the abnormal data of the purchase amount, and obtaining the processed data.

S2: Compare the original prediction model and the recurrent neural network model by a preset method;

Wherein, the S2 specifically includes the following steps: since the original prediction model cannot focus on long-term prediction friendly, therefore, in this embodiment, the original prediction model and the cyclic neural network model are analyzed from data processing, functional relationship, time dependence, The four aspects of the trend are compared and analyzed, and a more suitable prediction model is found according to the defects of the original prediction model. Table 1 shows the comparison between the original model and the recurrent neural network model.

Table 1 Comparison of original model and recurrent neural network model

LSTM (Long Short-Term Memory) is a long short-term memory network, a time recurrent neural network, suitable for processing and predicting important events with relatively long intervals and delays in time series, as shown in Figure 4 is a typical LSTM model Structure, in the figure x _t represents the input at time t, h _t represents the output of the cell at this time, and the middle cell shows such a process of specific value transfer calculation.

There are three main stages inside LSTM:

1) Forgetting stage: This stage is mainly to selectively forget the input from the previous node. In short, it will "forget the unimportant, remember the important", and use the calculated f _t+1 as the forget gate to control which c _t-1 of the previous state need to be forgotten.

2) Select the memory stage: This stage selectively "memory" the input of this stage. The main thing is to memorize the input x _t , which are important to focus on recording, and which are not important, remember a little less, the current input content is represented by the previously calculated C, and the control signal of the selection gate is controlled by C _t .

3) Output stage: This stage will decide which will be regarded as the output of the current state. It is mainly controlled by O _t+1 , and the O _t obtained in the previous stage is also scaled (changed by a tanh activation function).

Specifically, as shown in S3, the construction of the LSTM model and the expression of the model are shown.

S3: using a preset method to construct an LSTM model, and using the processed data as the input of the LSTM model to obtain a predicted value;

Wherein, the S3 specifically includes the following steps:

S31: establish an LSTM model by a preset method;

Specifically, the S31 specifically includes the following steps:

S311: Selectively forget the input information from the previous node, and obtain the forgetting threshold f _t+1 as the forgetting gate by calculating, to control which c _t-1 of the previous state need to be forgotten, and the forgetting threshold f _t+1 The expression is:

f _t+1 =σ(W _f ·[h _t , x _t ]+b _f );

Among them, h _t represents the output of the unit at time t, W _f represents the weight of the forget gate, b _f represents the bias of the forget gate, x _t represents the input at time t, and σ represents the sigmod function;

S312: Selectively memorize the input information, use i _t+1 to represent the input threshold, and use the cell state C _t at time t to control the control signal of the selection gate, input the threshold it _{+ 1} and the cell state at time t The expressions of C _t are:

i _t+1 =σ(W _i ·[h _t , x _t ]+b _i ), C _t =tanh(W _c ·[h _t , x _t ]+b _c );

Among them, Wi represents the weight of the input gate, _bi represents the bias of the input gate, W _c represents the weight of the cell, _{and bc represents} the bias of the cell;

S313: Update the state C _t of the cell at time t to obtain the state C _t+1 of the cell at time t+1, and the expression of the state C _t+1 of the cell at time t+1 is:

C _t+1 =f _t+1 *C _t +i _t+1 *C _t ;

S314: Controlled by the output threshold O _t+1 , and scaling the O _t obtained in the previous stage to obtain the output h _{t+1 at time t+1} , the output thresholds O _t+1 and the output at time t+1 The expressions of h _t+1 are:

O _t+1 =σ(W _o ·[h _t , x _t ]+b _o ), h _t+1 =O _t+1 *tanh(C _t+1 );

where W _o represents the output gate weight, and b _o represents the output gate bias.

S32: Use the processed data as the input sequence of the LSTM model, and mark it as {x ₁ , x ₂ ,...x _t ...}, where x _t represents the input at time t;

S33: Train the _LSTM model to obtain forget gate weight W _f , forget gate bias b _f , input gate weight Wi , input gate bias _bi , cell weight W _c , cell bias b _c , and output gate weight W _o and output gate bias b _o ;

Specifically, the traditional model prediction requires a large amount of data as support, but the prediction of the State Grid’s material procurement cannot fit the linear equation of the procurement volume well due to the insufficient data volume. In order to make the use of the existing data volume In this example, when training the model, the cycle is 1 day, 7 days, 8 days, 10 days, and 12 respectively (for example, if the cycle is 7 days, the data from No. 1 to No. 7 are added together as a Data, the data from No. 8 to No. 14 are added as one data, and the latter data are analogized) for training, and then the model with the best effect is selected for prediction;

Due to the uneven data distribution of each item, and in the training process, insufficient training and overfitting are often caused because the training data is too small or the dispersion is too high. Overfitting refers to the fact that due to too little training data, or too many training times for the training set, the result of the model is not to find the general common characteristics of all the data, but only to perform feature extraction on the training data. In other words, the model has memorized all the training data and predicts very well on the training data but very poorly on other data.

In the case of insufficient training, the training effect can be achieved by increasing the nodes in the network or increasing the training period of the network;

In the case of overfitting, the training period can be reduced or controlled, and the training of the network can be stopped before the inflection point of the data to achieve the training effect.

S34: After the LSTM model is trained, input the processed data at time t into the trained LSTM model, and obtain the data at time t+1, which is the predicted value.

Specifically, when training an LSTM model, it may be necessary to scale the data for sequence prediction problems. When the distribution of the input data sequence is not standard, or the variation range (standard deviation) is too large, it will slow down the learning and convergence speed of the network, and also hinder the learning efficiency of the network. Since there are outliers and more noise in the item purchase data set, normalization is used in this embodiment to reprocess the data at time t after processing, which can indirectly avoid the influence of outliers and extreme values through centralization, and normalize The formula is: x=(x-μ)/σ, where x represents the original value, μ represents the mean value, and σ represents the variance.

In order to better understand the technical solution, this implementation also includes two parts: error evaluation and prediction result display.

1) Error evaluation

In order to evaluate the prediction performance of the LSTM neural network model after the introduction of the new feature group on the original sequential time series, the mean absolute error (MAE), root mean square error (RMSE) and relative accuracy were selected for error evaluation. Because at the time point when the purchase quantity is too small, when evaluating the relative accuracy, the error will be much larger than at the time point when the purchase quantity is relatively large, and the purchase quantity is too small and there is no need to predict, so delete it when calculating the relative accuracy rate The data is too different from the actual average purchase volume.

MAE能更好的反映预测值与真实值误差的实际情况，MAE值越大，表明预测值与真实值之间的差别越大，预测效果越差，反之越好。①Mean absolute error (MAE), refers to the average value of the absolute value of the deviation between all predicted values and the true value, and its expression is:

MAE can better reflect the actual situation of the error between the predicted value and the actual value. The larger the MAE value, the greater the difference between the predicted value and the actual value, the worse the prediction effect, and vice versa.

RMSE对于一组预测的特小或特大误差非常敏感，可以很好反映预测的精密度，因此同MAE一样，RMSE越大，预测效果则越差。②The root mean square error (RMSE) refers to the average value of the square of the difference between all predicted values and the actual value, and then the value obtained by the square root, its expression is;

RMSE is very sensitive to very small or large errors in a set of predictions, and can well reflect the precision of prediction. Therefore, like MAE, the larger the RMSE, the worse the prediction effect.

③ Relative accuracy, relative accuracy is different from MAE and RMSE. The higher the accuracy, the better the prediction effect and the higher the prediction accuracy. The expression is: relative

In the above expression, f _i represents the predicted value of point i, y _i represents the actual value of point i, and N represents the number of data.

2) Prediction result display

Establish a procurement forecasting system based on the procurement requirements of goods, carry out the whole process modeling of the procurement chain, promote the construction of a unified data system for the entire supply chain of power materials, develop a set of preprocessing algorithms suitable for future procurement requirements of power materials, and use the procurement demand forecast model. Verify and guide the collection of traditional procurement requirements to the final step-by-step replacement and IT solidification, so that the management of procurement requirements of telecom operators is scientific, refined and efficient, freeing manpower and improving efficiency in the procurement process, accurate forecasting and lean management of procurement requirements. , Purchasing in a planned way, effectively reducing unexpected procurement costs, warehousing costs and engineering time costs, and achieving procurement optimization and cost optimization.

In order to verify the improvement of the accuracy of the model after the introduction, 6 categories with higher procurement frequency are selected for prediction in this implementation. They are 10kv transformers, 10kv cable terminals, low-voltage switch cabinets, power cables, overhead insulated conductors, and grounding lines. In this implementation, the above 6 types of data are input into the previously written prediction model for training. Table 2 shows the data display of the absolute average error, root mean square error and accuracy rate after each type of training (taking the data of different periods as Input, select the model with the training results).

Table 2 Data of each category after training by LSTM model

During the training of item purchase prediction, it was found that the results of training in cycles of 1 day, 7 days, 8 days, 10 days, and 12 were completely different, because the data distribution of each item for different cycles was different, and the discreteness was also inconsistent. , so this example uses different prediction periods for different items. Taking a 10kv transformer as an example, the results of training with different time periods are shown. It can be seen that the absolute squared difference, root mean square error and accuracy rate of The difference, as shown in Table 3, is the result after the data of the 10kv transformer is input to the prediction model training.

Table 310kv transformer training results

Time period 1 day 7 days 8 days 10 days 12 days MAE 12.58 10.98 10.34 15.85 11.55 RMSE 15.48 20.65 19.43 18.23 27.79 Accuracy 0.54 0.63 0.72 0.93 0.74 Time (s) 959.05482 892.88947 883.89589 861.86243 743.85834

Among them, Figure 5-10 shows the prediction results of 6 categories of LSTM model. In the process of training each category, we tried to train in cycles of 1 day, 7 days, 8 days, 10 days and 12 days. The distribution of the data is different, and the degree of dispersion is also different. During the training process, it was found that the 10kv transformer and the grounding line were trained in a period of 10 days. The results are better, and can achieve better prediction accuracy. 10kv cable terminal . The overhead insulated conductor is trained in a cycle of one week, which can achieve good prediction accuracy. The low-voltage switchgear is trained in a cycle of 8 days, and the power cable is trained in a cycle of 12 days, which can achieve their respective optimal predictions. precision.

According to the prediction results in this example, it is found that the prediction accuracy of only the 10kv transformer and the grounding line has reached more than 90%. Because of their large amount of data, good data distribution regularity and low discreteness, the time feature group can be well extracted. , so it can achieve good results in prediction training.

Specifically, as shown in Figure 5, the training results of the 10kv transformer are shown. The solid line is the real value, and the dotted line is the predicted value. Since the data volume of the 10kv transformer is large and the dispersion is not high, the prediction model can extract the Characteristic, the prediction period of the 10kv transformer is 10 days. It can be clearly seen from the figure that the prediction accuracy of the 10kv transformer is very high, reaching 0.93.

Figure 6 shows the prediction results of the 10kv cable terminal, the solid line is the real value, the dotted line is the predicted value, the prediction period of the 10kv cable terminal is 7 days, the original data of the 10kv cable terminal is too discrete, and there is still too high after data processing Too low data, resulting in the final prediction accuracy is not very high, reaching 0.78.

Figure 7 shows the prediction result of the low-voltage switchgear. The solid line is the actual value, and the dotted line is the predicted value. The prediction period of the low-voltage switchgear is 8 days, and the prediction accuracy of the low-voltage switchgear is not high, only 0.69.

Figure 8 shows the prediction result of the power cable, the solid line is the real value, the dotted line is the predicted value, the prediction period of the power cable is 12 days, and the prediction accuracy of the power cable is 0.69.

Figure 9 shows the prediction result of the overhead insulated wire, the solid line is the real value, the dotted line is the predicted value, the prediction period of the overhead insulated wire is 7 days, and the prediction accuracy is 0.8.

As shown in Figure 10, the prediction result of the grounding line is the real value, the dashed line is the predicted value, and the forecasting period of the grounding line is 10 days. Due to the large amount of data of the grounding line and the low degree of dispersion, it can be extracted very well. Among them, the prediction accuracy reached 0.9.

The prediction principle of this scheme is as follows: Since there are many influencing factors on the purchasing quantity of materials, to find these influencing factors, it is necessary to establish a model of influencing factors and purchasing quantity, and fit the functional relationship between the influencing factors and purchasing quantity. However, there are too many influencing factors of material purchases, and ordinary linear equations cannot completely fit the relationship between purchases and influencing factors, and the purchase of materials by the State Grid is closely related to the time series, while the LSTM network The method proposed by the model can just fit a nonlinear functional relationship, and can make good use of (how to use it in the data input in the model) data with time series, showing some relationships of time series data .

To sum up, with the help of the above technical solutions of the present invention, the data is processed through data cleaning, and the abnormally high and low purchase volume is removed, so that the data has low discreteness and certain regularity, thereby effectively avoiding abnormal values. The impact on the prediction has improved its prediction accuracy.

In addition, the present invention constructs the LSTM model by comparing the original prediction model and the cyclic neural network model, and can better fit the complex joint distribution through the nonlinear relationship, and map the relationship between the input and output well, so that the prediction model can be flexibly applied It depends on the distribution of the responsible joints to ensure its prediction effect.

In addition, by constructing and training a material procurement prediction model based on an LSTM network, the present invention considers the time series relationship between historical procurement quantities, increases or decreases the level according to the time series, and can repeat the pattern over time, improving the model's performance. prediction accuracy.

In addition, the present invention uses the updated cell state at time t-1, the output of the cell, and the data at time t as input, and through the combined use of three inputs and one output, the LSTM model can remember the long-term state, so that the It can effectively solve the long-distance dependence problem.

In addition, the present invention verifies and guides the collection of traditional procurement requirements through the procurement demand prediction model, and guides the final step-by-step replacement and IT solidification, so that the procurement demand management work of telecom operators is scientific, refined and efficient, and the procurement process releases manpower and improves efficiency. ,Accurate prediction of procurement requirements, lean management, and planned procurement, effectively reduce sudden procurement costs, warehousing costs and engineering time costs, and achieve procurement optimization and cost optimization.

In the present invention, the above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be Included in the protection scope of the present invention.

Claims (8)

1. a method based on LSTM network prediction material procurement demand, is characterized in that, comprises the following steps: S1: use a preset method to perform data processing on the pre-collected data; S2: Compare the original prediction model and the recurrent neural network model by a preset method; S3: Construct an LSTM model by using a preset method, and use the processed data as the input of the LSTM model to obtain a predicted value. 2. a kind of method based on LSTM network prediction material procurement demand according to claim 1, is characterized in that, described step S1 adopts preset method to carry out data processing to pre-collected data specifically comprises the following steps: S11: Collect the required data through the prepared database to obtain the collected data; S12: Perform data cleaning on the collected data by a preset method to obtain processed data. 3. A method for predicting material procurement demand based on LSTM network according to claim 2, wherein the step S12 performs data cleaning on the collected data by a preset method, and the processed data specifically includes: The following steps: S121: Perform random sampling on the original time series data in the collected data and obtain a plurality of pieces of estimated data, and fill up the gaps and defects generated by the random sampling to obtain a plurality of pieces of fill-in estimated data; S122: Classify all the supplemented estimated data according to the sampling time point, obtain multiple groups of time-classified data, and sort each group of the time-classified data according to size to obtain multiple groups of sorted arrays; S123: Process multiple groups of the sorted arrays to obtain a plurality of corresponding average data, and form an average sequence by using the plurality of average data; S124: Outputting the mean value sequence, that is, completing the removal of the abnormal data of the purchase amount, and obtaining the processed data. 4. a kind of method based on LSTM network prediction material purchase demand according to claim 1, is characterized in that, described step S2 is by preset method to compare original prediction model and cyclic neural network model specifically comprises the following steps: The original prediction model and the recurrent neural network model were compared and analyzed from four aspects: data processing, functional relationship, time dependence, and trend, and a more suitable prediction model was found according to the defects of the original prediction model. 5. a kind of method based on LSTM network prediction material procurement demand according to claim 1, is characterized in that, described step S3 adopts preset method to construct LSTM model, and described processed data is used as described LSTM The input of the model to obtain the predicted value includes the following steps: S31: establish an LSTM model by a preset method; S32: Use the processed data as the input sequence of the LSTM model, and mark it as {x ₁ , x ₂ ,...x _t ...}, where x _t represents the input at time t; S33: Train the _LSTM model to obtain forget gate weight W _f , forget gate bias b _f , input gate weight Wi , input gate bias _bi , cell weight W _c , cell bias b _c , and output gate weight W _o and output gate bias b _o ; S34: After the LSTM model is trained, input the processed data at time t into the trained LSTM model, and obtain the data at time t+1, which is the predicted value. 6. a kind of method based on LSTM network prediction material procurement demand according to claim 5, is characterized in that, described step S31 establishes LSTM model by preset method specifically comprises the following steps: S311: Selectively forget the input information from the previous node, and obtain the forgetting threshold f _t+1 as the forgetting gate by calculating, to control which c _t-1 of the previous state need to be forgotten, and the forgetting threshold f _t+1 The expression is: f _t+1 =σ(W _f ·[h _t , x _t ]+b _f ); Among them, h _t represents the output of the unit at time t, W _f represents the weight of the forget gate, b _f represents the bias of the forget gate, x _t represents the input at time t, and σ represents the sigmod function;

S312: Selectively memorize the input information, use i _t+1 to represent the input threshold, and use the cell state C _t at time t to control the control signal of the selection gate, input the threshold it _{+ 1} and the cell state at time t The expressions of C _t are: i _t+1 =σ(W _i ·[h _t , x _t ]+b _i ), C _t =tanh(W _c ·[h _t , x _t ]+b _c ); Among them, Wi represents the weight of the input gate, _bi represents the bias of the input gate, W _c represents the weight of the cell, _{and bc represents} the bias of the cell; S313: Update the state C _t of the cell at time t to obtain the state C _t+1 of the cell at time t+1, and the expression of the state C _t+1 of the cell at time t+1 is: C _t+1 =f _t+1 *C _t +i _t+1 *C _t ; S314: Controlled by the output threshold O _t+1 , and scaling the O _t obtained in the previous stage to obtain the output h _{t+1 at time t+1} , the output thresholds O _t+1 and the output at time t+1 The expressions of h _t+1 are: O _t+1 =σ(W _o ·[h _t , x _t ]+b _o ), h _t+1 =O _t+1 *tanh(C _t+1 ); where W _o represents the output gate weight, and b _o represents the output gate bias. 7. a kind of method based on LSTM network prediction material procurement demand according to claim 5 is characterized in that, in described step S33, the training of described LSTM model further comprises the following steps: when training described LSTM model , training in cycles of 1 day, 7 days, 8 days, 10 days, and 12 days, and then select the model with the best training effect for prediction. 8. a kind of method based on LSTM network prediction material procurement demand according to claim 5, is characterized in that, in described step S33, the training of described LSTM model further comprises the following steps: In the case of insufficient training, the training effect can be achieved by increasing the nodes in the network or increasing the training period of the network; In the case of overfitting, the training period can be reduced or controlled, and the training of the network is stopped before the inflection point of the data to achieve the training effect.

Images