A method for forecasting short-term electrical load in steel plants

Abstract

The present invention relates to a technology for processing electric load data, and in particular to a method for predicting short-term electric load in a steel plant. The prediction method includes the following steps: obtaining three time series data; calculating the Spearman correlation coefficient; judging the correlation; constructing a time index increment sequence; normalizing the data; constructing an input layer; processing the input matrix; constructing an LSTM network; merging the outputs of each parallel LSTM network; and mapping the output of the LSTM to a single value, namely, the target time interval P. ft Energy consumption prediction. This patent proposes a LSTM-based framework to predict the short-term energy consumption of steel plants. Experimental results show that this method obtains the best prediction results on the actual data set.

Description

Technical Field

[0001] The invention relates to an electric load data processing technology, and in particular to a short-term electric load forecasting method for a steel plant. Background Art

[0002] Since electricity cannot be stored, power operators should ensure that a precise balance is maintained between electricity production and consumption at all times. Therefore, it is necessary to formulate precise dispatch plans, and accurate short-term load forecasting can significantly promote the operation of the power system. Short-term load forecasting mainly predicts the power load in the next few hours, one day or several days. In addition, the forecast error has a great impact on the subsequent safety inspection of the power grid. Short-term load forecasting is also of great significance for dynamic state estimation and load dispatching.

[0003] Smart grid is a hot topic in the current global power industry, and steel plants play an important role in smart grid demand response. Traditional short-term load forecasting models are divided into two categories. The first is the time series model, including the Box-Jenkins model, the autoregressive moving average model, and the autoregressive integrated moving average model. This forecasting method is simple to implement and widely used, but it has high requirements on the stability of the load series. However, compared with the system load, the energy consumption uncertainty of a single non-residential user is extremely high, and the load series of different types of users vary greatly. Therefore, when predicting the energy consumption of a single steel plant, the prediction performance of the time series model is usually poor. The second is to consider the strong correlation between load samples, including the two-stage forecasting method proposed by Penya et al. However, due to the high uncertainty of energy consumption, it is difficult to describe the energy consumption pattern of a single steel plant using only the day of the week type. Kong et al. proposed a short-term load forecasting framework based on long short-term memory (LSTM) for all residential users. However, current research only uses time series forecasting methods, and does not specifically extract the intrinsic time series features of the sequence and further construct a load forecasting model based on the time series features. The energy consumption of steel plants accounts for a large proportion of the total energy consumption and is closely related to the stable operation of the power grid. Therefore, it is very important to solve the problem of short-term steel plant load forecasting.

[0004] The first algorithm lacks consideration of the correlation between time series data. These methods only establish nonlinear mapping relationships between features and loads, while ignoring the correlation between continuous load samples. In fact, as a typical time series, load samples are not only nonlinear and non-Gaussian with their corresponding features, but also highly correlated with each other. For a given user, the change of load data is a continuous process. The load at the current moment and the load at the previous moment are not independent of each other, because the current load depends not only on the current characteristics, but also on the characteristics and loads at past moments. Therefore, most existing methods only establish nonlinear relationships between the features at a single moment and the corresponding load. They ignore the strong correlation between load samples, resulting in limited prediction accuracy.

[0005] The second algorithm only uses the time series prediction method, and does not specifically extract the intrinsic time series characteristics of the sequence and further build a load forecasting model based on the time series characteristics. Summary of the invention

[0006] The present invention proposes a method for predicting short-term electric load of a steel plant based on a long short-term memory (LSTM) recursive neural network nonlinear time series method based on time series characteristics, and utilizes the extracted time series information of multiple related sequences to improve the accuracy of LSTM method in predicting the load of the steel plant.

[0007] Explanation of relevant terms:

[0008] Long Short-Term Memory (LSTM) Recurrent Neural Network: Long Short-Term Memory (LSTM) is a special recurrent neural network (RNN) architecture designed to solve the long-term dependency problem encountered by standard recurrent neural networks when processing sequence data. LSTM effectively controls and transmits information by introducing "cell states" and three different gate structures: input gate, forget gate, and output gate. These gates can learn when to let information pass and when to ignore information, so that they can remember long-term sequence information and model the relationship between elements in the sequence with very long time intervals.

[0009] Self-organizing neural network SOM: Self-organizing neural network is an unsupervised learning algorithm mainly used for dimensionality reduction and visualization of high-dimensional data. SOM can map high-dimensional input data to a low-dimensional discrete space, usually two-dimensional, while maintaining the topological structure of the input data. In this low-dimensional space, adjacent neurons represent similar features in the input space. SOM is implemented through a competitive learning process, in which the "winning" neuron - the neuron that best matches the input sample and its neighboring neurons adjust their weights to better represent the input data.

[0010] Spearman correlation coefficient: Spearman correlation coefficient is a non-parametric rank correlation coefficient used to measure the dependence of two variables. It evaluates the correlation between the ranking of two variables rather than their specific values. The value range of Spearman correlation coefficient is between -1 and 1, where 1 indicates a perfect positive correlation, -1 indicates a perfect negative correlation, and 0 indicates no correlation. Unlike the Pearson correlation coefficient, the Spearman correlation coefficient does not require a linear relationship between variables or a normal distribution, so it is suitable for variables with nonlinear relationships or asymmetric distribution data.

[0011] Recurrent Neural Network (RNN) is a deep neural network that is good at processing time series data. Unlike traditional feedforward neural networks, RNN saves, learns, and records historical information in sequence data by connecting periodic hidden layer nodes. In order to solve the long-term dependency of RNN, that is, the gradient disappearance problem, LSTM came into being. Compared with RNN, LSTM can adaptively control the amount of information passing through these gates. In addition, LSTM is based on the idea of ​​creating paths in time, which has the advantage of neither disappearing nor exploding. These cells accumulate information over a long period of time, such as evidence of specific features or categories. Therefore, LSTM is good at capturing long-distance dependencies between sequence data and is widely used in tasks such as natural language processing and speech recognition.

[0012] A data set contains load data of 10 steel plants collected from a certain region in China from January 1 to December 31, 2015, with a sampling interval of 15 minutes. Figure 1-Figure 3 The hourly load curves of the three steel plants are shown. It can be seen that the two periods from 6:00 to 9:00 and from 18:00 to 20:00 are the peak periods of energy consumption. In addition, the energy consumption in the evening hours is higher than that in the morning hours. After 22:00, the energy consumption of the steel plants drops sharply and remains stable at around 1MW because the workers go off work. Since each steel plant uses electricity differently, its energy consumption pattern is completely different. Therefore, it is reasonable to train a prediction model for each steel plant separately.

[0013] Cluster analysis: For a specific steel plant, its production and management usually have strong periodicity and regularity. If there are no major events or policy changes, we assume that the load curve of a specific steel plant should have the following characteristics:

[0014] The load curve fluctuation in the short term should be small, and the correlation between adjacent time points should be strong.

[0015] There is no significant difference in load at the same time point on different days, and the energy consumption at different times of each day is highly correlated.

[0016] The energy consumption load curves of the same day in different weeks are very similar, and the weekly information is highly correlated.

[0017] The difference in daily load profile between working days and non-working days depends on the operation pattern of the steel plant industry.

[0018] To verify our hypothesis, the SOM algorithm is used to explore the pattern of energy consumption in the steel plant, such as Figure 4 As shown in Figure 1, the clustering results of the daily load curve of the steel plant in each month of the year. Since we assume that the activities of the steel plant can be divided into two types, namely working days or non-working days, in the experiment, it can be seen that most working days and non-working days are not in the same cluster. Based on the cumulative energy consumption data of each day in each month of the year, the similarity of energy consumption patterns is visualized through SOM cluster analysis, and the results are shown in Figure 1. Figure 4 shown.

[0019] From the two-dimensional SOM representation of data similarity, Figure 4 From the neighbor representation of the cells in the figure, it can be seen that {Friday, Saturday, Sunday} are relatively high energy consumption days in most months, i.e., they appear in clusters with higher color intensity red. However, slight differences between some months are obvious, for example, in February, in the fifth cluster with the highest energy consumption, {Sunday, Tuesday, Thursday} are the times with the highest energy consumption. In addition, for the highest energy consumption cluster II, {Saturday, Sunday, Monday} are the times with the highest energy consumption, with an occurrence frequency of 75%. For June, the highest energy consumption is {Friday, Saturday}, all in the fourth cluster. However, the highest energy consumption in June is {Monday, Wednesday, Thursday}, which appears in the sixth cluster with a frequency of 100%. There are other exceptions, such as in July, the highest energy consumption is Thursday, with an occurrence rate of 30%, while in May, the highest energy consumption is on Sunday, which also occurs 50%. Interestingly, in most months, Tuesdays and Wednesdays usually appear in clusters with low energy consumption. Figure 4 The cluster analysis shown in also shows high energy consumption from June to August. At the same time, {Friday, Saturday, Sunday} often appears in the highest energy consumption cluster of {June, July, August} - the red cluster.

[0020] Time series correlation: Since the load data of steel plants do not come from a normal distribution, the Spearman correlation coefficient is used to measure the time series correlation of a single steel plant. Since the Spearman correlation coefficient eliminates some interference factors that affect the overall data, it is more applicable when the amount of data is large. Therefore, the present invention uses the Spearman coefficient in load forecasting to calculate the time series correlation, including the correlation of adjacent time points, daily correlation, and weekly correlation.

[0021] for Figure 4For the steel plant analyzed in , we obtained the Spearman correlation coefficients for the correlations at adjacent time points, daily correlations, and weekly correlations. 0.01,10 The T value calculated in the case of 0.01,10 , then accept hypothesis H1, X and Y are correlated. Otherwise, H0: X and Y are not correlated. The calculation results are shown in Table 1.

[0022] Table 1 Spearman correlation coefficients of adjacent time point correlations, daily correlations, and weekly correlations

[0023]

[0024] From the Spearman correlation coefficient in Table 1, it can be seen that there is a high correlation between Steel Plant 1 and Steel Plant 2 in terms of adjacent time point correlation, daily correlation correlation and weekly correlation correlation, all of which meet the H1 hypothesis. However, the Spearman correlation coefficient of Steel Plant 3 is low, and there is no obvious correlation in the adjacent time point series, daily correlation series and weekly correlation series.

[0025] In order to exploit different serial correlations and capture long-range dependencies in load data, we propose a steel plant load forecasting framework based on LSTM and vine copula, which takes each time series as input. Figure 5 As shown. There is time series correlation in the load data of most steel plants, including adjacent time point correlation, daily correlation and weekly correlation. Therefore, three LSTM recurrent neural networks are parallelized to capture the three kinds of related information in the time series. In addition, LSTM learns the energy consumption pattern of the steel plant through three time series, thereby improving the prediction accuracy. Under the LSTM framework, a short-term electricity load forecasting method for steel plants includes the following process:

[0026] S100: Obtain three time series data, including:

[0027] S110: Acquire the sequence data of adjacent time points; set the time step K1, and construct the load sequence E1, which contains the load data of K1 consecutive time points before the prediction time point;

[0028] S120: Acquire daily time series data; set the time step K2, and construct a load sequence E2, which contains the load data at the same time point within K2 consecutive days before the prediction time point;

[0029] S130: Obtain weekly time series data; set the time step K3, and construct a load sequence E3, which contains the load data of the same day and the same time point within K3 consecutive weeks before the prediction time point;

[0030] S200: Spearman correlation coefficient calculation; respectively calculate the Spearman correlation coefficient T1 of adjacent time point data, the Spearman correlation coefficient T2 of adjacent daily data, and the Spearman correlation coefficient T3 of adjacent weekly data;

[0031] S300: Correlation judgment, if T1>t 0.01,10 , it means that the load data at consecutive time points are related; if T2>t 0.01,10 , it means that the load data of consecutive days are related; if T3>t 0.01,10 , it means that the load data of consecutive weeks are related;

[0032] S400: construct a time index increment sequence; specific process:

[0033] S410: Create a sequence I∈R k , contains the time index increments for the past K1 time steps;

[0034] S420: Create a sequence D, representing the weekly index of the past K2 time steps;

[0035] S430: Create a binary sequence H to represent a specific time attribute, a weekday or a weekend;

[0036] S500: Data normalization; if T1>t 0.01,10 , and T2≤t 0.01,10 , and T3≤t 0.01,10 , go to S510; if T1>t 0.01,10 , and T2>t 0.01,10 , and T3≤t 0.01,10 , go to S520; if T1>t 0.01,10 , and T2≤t 0.01,10 , and T3>t 0.01,10 , go to S530; if T1>t 0.01,10 , and T2>t 0.01,10 , and T3>t 0.01,10 , transfer to S540;

[0037] S510: Use the minimum-maximum method to normalize the load sequence E={E1} so that its range is [0,1]; then go to S550;

[0038] S520: Use the minimum-maximum method to normalize the load sequence E={E1, E2} so that its range is [0, 1]; then go to S550;

[0039] S530: Use the minimum-maximum method to normalize the load sequence E={E1, E3} so that its range is [0, 1]; then go to S550;

[0040] S540: Use the minimum-maximum method to normalize the load sequence E={E1, E2, E3} so that its range is [0, 1]; then go to S550;

[0041] S550: Perform one-hot encoding on the time index increment sequence I, the week index sequence D, and the binary mark sequence H;

[0042] S600: construct an input layer; combine the normalized load sequence E, the one-hot encoded I, D, and H into an input matrix as the input of the LSTM model;

[0043] S700: Input matrix processing; each row of the input matrix represents the features of a time step, and these features are fed into the corresponding LSTM network;

[0044] S800: LSTM network construction; each sequence vector is fed into an independent LSTM network; each LSTM network is responsible for learning the time dependency of the corresponding sequence;

[0045] S900: Merge the outputs of each parallel LSTM network; the output layer maps the output of the LSTM to a single value, namely the target time interval P ft energy consumption forecast.

[0046] Preferably, in S110, adjacent time points are spaced 15 minutes apart; in S410, the element range is from 1 to 96, and the load data at each time point corresponds to a time sequence point; in S420, the range is from 0 to 6, corresponding to Monday to Sunday.

[0047] This patent proposes a LSTM-based framework to predict the short-term energy consumption of steel plants. The framework uses serial correlation features as input; for a specific steel plant, if the Spearman correlation coefficient of the steel plant is high and the hypothesis H1 is accepted, such time series features can be added to the proposed framework to improve the prediction performance. Experimental results show that this method obtains the best prediction results on the actual data set. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is the daily load curve of the energy consumption of the steel plant 1 of the present invention. The horizontal axis corresponds to the 24 hourly moments in a day, and the vertical axis represents the load in MW. In the figure, seven color curves represent different daily load curves for seven days of a week.

[0049] Figure 2It is the daily load curve of the energy consumption of the steel plant 2 of the present invention. The horizontal axis corresponds to the 24 hourly moments in a day, and the vertical axis represents the load in MW. In the figure, seven color curves represent different daily load curves for seven days of a week.

[0050] Figure 3 It is the daily load curve of the energy consumption of the steel plant 3 of the present invention. The horizontal axis corresponds to the 24 hourly moments in a day, and the vertical axis represents the load in MW. In the figure, seven color curves represent different daily load curves for seven days of a week.

[0051] Figure 4 It is the clustering result of the daily load curve of a steel plant in each month of the year; the red end represents high energy consumption, the blue end represents low energy consumption, and the colors between the two, from red to blue, indicate decreasing energy consumption.

[0052] Figure 5 This is a short-term steel plant load forecasting framework diagram based on LSTM of the present invention.

[0053] Figure 6 It is a comparison chart of MAPE values ​​of prediction results based on different time series combinations as input. DETAILED DESCRIPTION

[0054] A method for predicting short-term electric load of a steel plant includes the following steps:

[0055] S100: Obtain three time series data, including:

[0056] S110: Acquire the data of the adjacent time points, with the interval between adjacent time points being 15 minutes; set the time step K1, and construct the load sequence E1, which contains the load data of K1 consecutive time points before the prediction time point;

[0057] S120: Acquire daily time series data; set the time step K2, and construct a load sequence E2, which contains load data at the same time point within K2 consecutive days before the prediction time point;

[0058] S130: Obtain weekly time series data; set the time step K3, and construct a load sequence E3, which contains the load data of the same day and the same time point within K3 consecutive weeks before the prediction time point;

[0059] S200: Spearman correlation coefficient calculation; respectively calculate the Spearman correlation coefficient T1 of the data at adjacent time points, the Spearman correlation coefficient T2 of the data at adjacent days, and the Spearman correlation coefficient T3 of the data at adjacent weeks;

[0060] S300: Correlation judgment, if T1>t 0.01,10, it means that the load data at consecutive time points are related; if T2>t 0.01,10 , it means that the load data of consecutive days are related; if T3>t 0.01,10 , it means that the load data of consecutive weeks are related;

[0061] S400: construct a time index increment sequence; specific process:

[0062] S410: Create a sequence I∈R k , contains the time index increments of the past K1 time steps, the element range is from 1 to 96, and the load data at each time point corresponds to a time series point;

[0063] S420: Create a sequence D, representing the week index of the past K2 time steps, ranging from 0 to 6, corresponding to Monday to Sunday;

[0064] S430: Create a binary sequence H to represent a specific time attribute, a weekday or a weekend;

[0065] S500: Data normalization; if T1>t 0.01,10 , and T2≤t 0.01,10 , and T3≤t 0.01,10 , go to S510; if T1>t 0.01,10 , and T2>t 0.01,10 , and T3≤t 0.01,10 , go to S520; if T1>t 0.01,10 , and T2≤t 0.01,10 , and T3>t 0.01,10 , go to S530; if T1>t 0.01,10 , and T2>t 0.01,10 , and T3>t 0.01,10 , transfer to S540;

[0066] S510: Use the minimum-maximum method to normalize the load sequence E={E1} so that its range is [0,1]; then go to S550;

[0067] S520: Use the minimum-maximum method to normalize the load sequence E={E1, E2} so that its range is [0, 1]; then go to S550;

[0068] S530: Use the minimum-maximum method to normalize the load sequence E={E1, E3} so that its range is [0, 1]; then go to S550;

[0069] S540: Use the minimum-maximum method to normalize the load sequence E={E1, E2, E3} so that its range is [0, 1]; then go to S550;

[0070] S550: One-hot encoding is performed on the time index increment sequence I, the week index sequence D, and the binary label sequence H; the one-hot encoder maps the original elements from the classification feature vector with M cardinality to a new vector with M elements, where only the corresponding new elements are assigned 1, and the remaining new elements are assigned 0;

[0071] S600: construct an input layer; combine the normalized load sequence E, the one-hot encoded I, D, and H into an input matrix as the input of the LSTM model;

[0072] S700: Input matrix processing; each row of the input matrix represents the features of a time step, and these features are fed into the corresponding LSTM network;

[0073] S800: LSTM network construction; each sequence vector is fed into an independent LSTM network; each LSTM network is responsible for learning the time dependency of the corresponding sequence;

[0074] S900: Merge the outputs of each parallel LSTM network; the output layer maps the output of the LSTM to a single value, namely the target time interval P ft energy consumption forecast.

[0075] Steel Plant Energy Consumption Model Performance Evaluation:

[0076] Single-plant prediction was carried out on the data set of 10 steel plants, and finally the average prediction mean absolute error MAE (unit: KW); root mean square error RMSE (unit: KW); mean absolute percentage error MAPE (unit: %) of each steel plant were taken as the final experimental result verification indicators.

[0077] Table 2 gives the prediction performance of the different models. In general, the prediction for each time interval for each steel plant is not as accurate as the prediction for the substation load. This level of accuracy is expected, as the prediction accuracy tends to drop significantly with decreasing aggregation levels. Therefore, the accuracy of the various methods in Table 2 is reasonable.

[0078] Table 2 Prediction performance of different comparison models

[0079]

[0080] As can be seen from Table 2, the LSTM model prediction accuracy has the best prediction performance under each statistical indicator. At the same time, the framework based on multi-sequence LSTM is generally the best. The best prediction performance can be obtained when the LSTM time step associated with the time point is set to 5. The prediction performance of RF is second only to LSTM, which is because the integrated learning has good generalization performance. The traditional ARIMA prediction MAPE is 35.87%, which is slightly better than SVR and ELM. The prediction performance of SVR and ELM belongs to the same level, and their prediction MAPE ranges from 70% to 90%. As the time step increases, the MAPE of ELM increases, which may be because ELM cannot utilize sequence information. When the time step increases, ELM regards it as noise and the performance decreases.

[0081] Model prediction performance under different circumstances: There are multiple time-related sequences in the data sample of a specific steel plant, such as adjacent time points, denoted as E1; the same time points of adjacent days, denoted as E2; the same day of adjacent weeks, denoted as E3. Therefore, we formed four comparative experiments with different characteristics.

[0082] LSTM model based on single temporal relationship: E1 is predicted based on an LSTM framework that only uses the relevant information of adjacent time points.

[0083] LSTM model based on partial temporal relations: E1 and E2 are predicted by the LSTM-based framework, which uses the information related to adjacent time points and the information related to the day; E1 and E3 are predicted by the LSTM-based framework, which uses the information related to adjacent time points and the information related to the week.

[0084] Multi-time series based LSTM model: E1, E2, and E3 are predicted using an LSTM-based framework that exploits all three sequence information.

[0085] These four experiments are designed to verify the contribution of different features to the model. Similarly, we conducted single steel plant prediction on the dataset of 10 steel plants and averaged the predicted MAPE of each steel plant in the final experimental results. The LSTM time point related time step is set to 5, as shown in Figure 6 shown.

[0086] from Figure 6 It can be seen that as the number of features increases, MAPE gradually increases. In the LSTM-based framework, the week-related features contribute more than the day-related features. In addition, the prediction results of the model that uses the three related features E1, E2, and E3 are the best among all the results.

[0087] From the perspective of load forecasting for a single steel plant, a steel plant is selected from each steel plant, and the MAPE values ​​are shown in Table 3. As can be seen from Table 3, the contributions of the day-related features and the week-related features are sometimes high and sometimes low. It can be inferred that this depends on the energy consumption pattern of each steel plant as analyzed in Section 3. In addition, most of the best results are in the model with three sequence features.

[0088] Table 3 MAPE values ​​of different characteristic matrices of a single steel plant

[0089] Steelplants <![CDATA[E1(%)]]> <![CDATA[E1E2(%)]]> <![CDATA[E1E3(%)]]> <![CDATA[E1E2E3(%)]]> Plant1 47.63 42.94 47.58 38.53 Plant2 11.76 11.74 11.61 11.52 Plant3 44.01 41.23 42.83 42.23 Plant4 30.00 23.38 17.68 16.93 Plant5 27.17 26.96 26.93 26.18 Plant6 47.07 46.01 43.02 40.75 Plant7 9.39 8.10 7.92 7.82 Plant8 57.56 55.82 53.82 54.70 Plant9 26.38 25.46 25.01 23.86 Plant9 19.04 18.70 18.69 18.52

[0090] Comparison of prediction performance of different steel plants: Figure 6 As can be seen in Figure 1, not all steel mills have the best prediction results using three features. At the same time, it also depends on the energy consumption pattern of the steel mill. If the pattern is irregular, the LSTM model cannot capture the time series correlation of the steel mill, and adding more time series features will reduce the prediction performance.

[0091] The proposed framework is used to predict and conduct four comparative experiments based on the load data characteristics of Steel Plant 1, Steel Plant 2, and Steel Plant 3. The MAPE values ​​of the prediction indicators of different steel plants based on different characteristics are shown in Table 4.

[0092] Table 4 Prediction indicators MAPE of different steel plants based on different characteristics

[0093] Steel Plant <![CDATA[E1(%)]]> <![CDATA[E1E2(%)]]> <![CDATA[E1E3(%)]]> <![CDATA[E1E2E3(%)]]> Steel Plant 1 17.72 16.04 16.10 15.45 Steel Plant 2 26.6 20.11 19.87 19.57 Steel Plant 3 17.00 18.57 17.66 19.5

[0094] As shown in Table 1, the Spearman correlation coefficients of Steel Plant 1 and Steel Plant 2 are high in the adjacent time point correlation, daily correlation, and weekly correlation, and both accept hypothesis H1. Therefore, in Table 4, the daily correlation features and weekly correlation features are added, and the prediction results become better. When the established model is based on all three feature vectors E1E2E3, the prediction performance is the best, and its prediction MAPE index is 2.27% and 7.03% lower than the MAPE index based on feature E1, respectively. However, when the LSTM model is used to predict the load data of Steel Plant 3, the LSTM model may not be able to learn effective information well, nor can it obtain better results, because the steel plant load data has no relevant information in the time series. The results are consistent with those shown in Table 1. The Spearman correlation coefficient of Steel Plant 3 is low in the adjacent time point correlation, daily correlation, and weekly correlation, and both accept hypothesis H0.

[0095] As mentioned earlier, the LSTM model can capture long-range dependencies in the sequence in order to discover the energy consumption pattern of a single steel plant. In addition, the Spearman correlation coefficient is a typical indicator of correlation estimation. For a specific steel plant, if the Spearman correlation coefficient of the steel plant is high and hypothesis 1 is accepted, such time series features can be added to the proposed framework to improve the prediction performance.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting short-term electricity load in a steel plant, characterized in that: The process includes the following: S100: Obtain three time series data, including: S110: Acquire the sequence data of adjacent time points; set the time step K1, and construct the load sequence E1, which contains the load data of K1 consecutive time points before the prediction time point; S120: Acquire daily time series data; set the time step K2, and construct a load sequence E2, which contains the load data at the same time point within K2 consecutive days before the prediction time point; S130: Obtain weekly time series data; set the time step K3, and construct a load sequence E3, which contains the load data of the same day and the same time point within K3 consecutive weeks before the prediction time point; S200: Spearman correlation coefficient calculation; respectively calculate the Spearman correlation coefficient T1 of the data at adjacent time points, the Spearman correlation coefficient T2 of the data at adjacent days, and the Spearman correlation coefficient T3 of the data at adjacent weeks; S300: Correlation judgment, if T1>t 0.01,10 , it means that the load data at consecutive time points are related; if T2>t 0.01,10 , it means that the load data of consecutive days are related; if T3>t 0.01,10 , it means that the load data of consecutive weeks are related; S400: construct a time index increment sequence; specific process: S410: Create a sequence I∈R k , contains the time index increments for the past K1 time steps; S420: Create a sequence D, representing the weekly index of the past K2 time steps; S430: Create a binary sequence H to represent a specific time attribute, a weekday or a weekend; S500: Data normalization; if T1>t 0.01,10 , and T2≤t 0.01,10 , and T3≤t 0.01,10 , go to S510; if T1>t 0.01,10 , and T2>t 0.01,10 , and T3≤t 0.01,10 , go to S520; if T1>t 0.01,10 , and T2≤t 0.01,10 , and T3>t 0.01,10 , go to S530; if T1>t 0.0110 , and T2>t 0.0110 , and T3>t 0.0110 , transfer to S540; S510: Use the minimum-maximum method to normalize the load sequence E={E1} so that its range is [0,1]; then go to S550; S520: Use the minimum-maximum method to normalize the load sequence E={E1, E2} so that its range is [0, 1]; then go to S550; S530: Use the minimum-maximum method to normalize the load sequence E={E1, E3} so that its range is [0, 1]; then go to S550; S540: Use the minimum-maximum method to normalize the load sequence E={E1, E2, E3} so that its range is [0, 1]; then go to S550; S550: Perform one-hot encoding on the time index increment sequence I, the week index sequence D, and the binary mark sequence H; S600: construct an input layer; combine the normalized load sequence E, the one-hot encoded I, D, and H into an input matrix as the input of the LSTM model; S700: Input matrix processing; each row of the input matrix represents the features of a time step, and these features are fed into the corresponding LSTM network; S800: LSTM network construction; each sequence vector is fed into a separate LSTM network; each LSTM network is responsible for learning the temporal dependency of the corresponding sequence; S900: Merge the outputs of each parallel LSTM network; the output layer maps the output of the LSTM to a single value, namely the target time interval P ft energy consumption forecast.

2. A method for predicting short-term electricity load in a steel plant according to claim 1, characterized in that: In the S110, adjacent time points are spaced 15 minutes apart; in S410, the element range is from 1 to 96, and the load data at each time point corresponds to a time sequence point; in S420, the range is from 0 to 6, corresponding to Monday to Sunday.

Images