Some New Developments of Forecasting

in Power Market
Hanlin Liu, Chengjin Ye, Yi Ding, Zhangyi Chen Xiang Xia, Jianliang Fang, Wei Jiang
Zhejiang University State Grid Zhejiang Electric Power Co., LTD.
Hangzhou, China Hangzhou, China
terence.liu@foxmail.com pxisx@163.com

Abstract—With the deregulation and development of grid, the focus of demand forecasting technology shifts from
power market, some new features and trends about electric a long period to a short period to adapt to the trading
power demand forecasting issues come into being, which behavior in the electricity market environment.
mainly includes the following two aspects: Firstly, the
forecasting has developed into a complicated problem In recent years, with the development of electricity
associated with users’ behavior and psychology. Secondly, market, how to adapt the demand forecasting techniques to
short and medium-term forecasting used for bidding and the electricity market has become an important and arduous
clearing in power market are discussed more than long-term task. And how to make full use of the data that is available
forecasting in order to guide power grid planning. from power system has become a topic needed to be
Considering the new features of forecasting and the great researched as soon as possible.
progress in artificial intelligence (AI), this paper puts forward
the idea and framework of forecasting problem under the
II. POWER MARKET CONSTRUCTION IN CHINA -
power market environment by integrated model and Deep
Learning technology. ZHEJIANG AND GUANGDONG AS EXAMPLES
China will establish a power market system with spot
Index Terms—Power Market; Demand Forecasting; market. The spot market can be divided into day-ahead
Artificial Intelligence transactions, intraday transactions and real-time transactions
according to trading time scales.
I. INTRODUCTION
The main function of day-ahead market is to provide a
Power system demand forecasting uses the information platform for bidding. Then, according to the demand supply
about power system, economy status, society situation and curve, the power plan of next day (usually hourly) is
meteorology to develop a pre-made estimation and determined. Intraday market makes a slight adjustment to
speculation about the demand in the power system. It the trading plan that has been set out. It is an important
analyze and study the historical data which tells the internal complement to the day-head market. The real-time market
relations between things[1]. adjusts the trading amount according to the quotation before
closing the system so as to ensure the instantaneous balance
At present, the mainstream research mode of demand
of system power.
forecasting is based on the time scale[2]. The long-term
forecast (more than 1 year) mainly provides suggestions for Fig.1 shows the organization of the spot market, and
the network planning. The medium-term forecast (more than Fig.2 shows the rolling correction in detail.
1 month) mainly provides reference for futures trading,
reservoir scheduling, overhaul and fuel plan. The Short-term
forecast (1 day - 1 week) is the basis of the spot transaction.
And the very short-term forecast (5 mins - 60 mins) is
mainly for real-time scheduling, real-time price forecast.
With the gradual evolution of electricity market, the
demand forecasting technology is changing at the same
time. Considering the electricity market, the first thing we
should take into account is the integration of various new
energy sources such as solar energy and wind energy. And
secondly, we should consider the transactional interactions
between supply and demand. The two points above will
bring some challenges to the traditional load forecasting
technology[3].
Fig 1. The Organization of Spot Market
Because of the slowdown of the construction of power
This work was supported in part by the science and technology project
of State Grid Zhejiang Electric Power Company, the National Key
Research and Development Program of China (2016YFB0901103).

978-1-5386-5186-5/18/$31.00 ©2018 IEEE

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on July 29,2020 at 09:49:29 UTC from IEEE Xplore. Restrictions apply.
 III. TRADITIONAL DEMAND FORECASTING TECHNOLOGY

A. Statistics based models

1) Regression analysis
The regression model is statistical process that find the
relationship between independent and dependent
variables[4]. After determining the correlation these
Fig 2. Detail of Rolling Correction variables, a specific mathematical model can be obtained.
The regression models can be presented as a math equation.
A. Overview of Zhejiang Electric Power Market
2) Exponential Smoothing
In January 2000, the construction of Zhejiang Power Exponential smoothing is a method of sequence
Market was completed, and the power market was put into analysis. The result that the method predicted is a weighted
operation at the same time. After three years of actual arithmetic average of historical data. The recent data has big
operation, it is the unique power generation market that is weight, while the data that was measured before has small
running in China. Thus, Zhejiang accumulated rich weight. Therefore, this model has an accurate result when
experience in the construction and operation of power predict the data that is near to the current time.
markets.
3) Random Time Series
In recent years, Zhejiang has made many explorations
and attempts during the reform of the electricity market. A The historical data of power load is recorded in a certain
series of changes have taken place in electricity transaction. sampling interval time series with a strong randomness. One
Initially, the type of electricity transaction is the effective method to deal with random sequences is the time
matchmaking transaction. Then, this platform became more series model, so it has a wide range of applications in short-
complex. Some new types of electricity transaction such as term prediction. However, this method doesn't care the
biding for clearance and bilateral trading has been factors that affect the load change. Thus, the accuracy or the
developed. forecasting is limited. Time series model is usually
combined with other methods in short-term forecasting to
In 2017, Zhejiang expanded the scale of direct improve the accuracy.
transactions furthermore. Nuclear power and various types
of incoming Energy from other provinces and cities are B. Artificial intelligence based modeling
included in direct transactions. Moreover, Business users
1) Support Vector Machine(SVM)
who consumed more than 1 million kWh of electricity each
SVM is based on the principles of statistical theory and
year participated in the trial as well. The number of power
structural risk minimization. It is developed from an
generation enterprises who involved in the pilot reached
artificial intelligence system, and the basic theory is to use
58,000. And the amount of the transaction power reached 93
the sample space to find the optimal hyperplane. The
billion KWh, which accounts for about one-quarter of the
hyperplane divides the sample set into several classes, and
total electricity consumption.
the distance between each class is maximized. SVM is
suitable for short-term load forecasting and can solve
B. Overview of Guangdong Electric Power Market problems with specific features, such as small samples,
Up to the end of 2016, 376 big users and 38 big coal- nonlinear and high-dimensional data. However, the training
fired power plants in Guangdong Province were allowed to time would be long when dealing with large sample data,
participate in the pilot of power market. The installed and the parameters selected in the model also affects the
capacity of these power plant is 50.9 million kilowatts, final prediction accuracy.
accounting for 53.4% of installed capacity in the province.
The bidding rules adopted in Guangdong are centralized 2) Artificial Neural Network (ANN)
matching trading algorithms which base on declaring the Artificial neural network is composed of multiple
spread. neurons, it simulates the way that human learn and process
the information[5]. Theoretically, artificial neural network
The algorithm matches the declaring spread for the can approximate to any function. ANN has some advantages
buyer and the supplier to form a declaring spread pair. The such as adaptive learning, function approximation and so on.
formulation of calculating the declaring spread pair is: It has turned out that the ANN model has achieved good
declaring spread pair = declaring spread from buyer results in short-term prediction. However, this model also
- declaring spread from supplier has some disadvantages since it lacks theoretical support
when initializing the network structure, and the training time
Once the result of the pair is negative, the trade between that the model takes is long when data is big, and the results
buyer and supplier will not be carried out, while the pair that the model gives may fall into local minima.
with positive or zero result will be carried out. The pair with
bigger spread will have higher priority. 3) Gray System
The gray system forecasting method uses partial explicit
Taking the transaction situation in January 2017 as an information to seek the correlation between various
example, the ratio between supply and demand of the parameters. The advantage of the gray forecasting model is
auction reached 2.24. That meant the supply side had a that it does not need to calculate the statistical characteristics
fierce competition, and the price reduction was beyond during the modeling. So, theoretically, it can be applied to
expectation. The average spears pair is -0.03894642 any non-linear demand forecasting. However, the
RMB/kWh. disadvantage is that the solution of the differential equation

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on July 29,2020 at 09:49:29 UTC from IEEE Xplore. Restrictions apply.
 is only suitable for the index that has an exponential growth A. The characteristics of demand forecasting in the new
trend. situation
4) Wavelet Analysis In economics, the spot market generally refers to real-
Wavelet analysis is one of the most outstanding time or the transactions near the real-time. In electricity
achievements of the 20th century mathematics research trading, the spot market refers to the day-ahead and intraday
results. Wavelet analysis is a time-frequency analysis transaction. The electricity spot market has a strong
method. It can divide the domain signals that mix with connection to the real-time operation of the electricity
different frequency into some block information on different market. Because of that, some new requirements are put
bands. For example, the demand signal can be decomposed forward to demand forecasting.
according to its own characteristics. Fig.4 shows some new requirements in forecasting.

C. Combination Method
Single forecasting model that uses only one method to
forecast the demand information is usually difficult to
predict accurately. Firstly, the load data is of complexity and
randomness. Secondly, each forecasting method is of
complexity and randomness, too. Thirdly, there are some
limitations in each forecasting method. Therefore,
researchers proposed a combined model to forecast. That is,
according to the advantages and disadvantages of various
algorithms, several algorithm models are reasonably
combined to achieve the purpose of improving the accuracy
of the prediction result.
While forecasting model that uses only one method to
forecast the demand information is usually difficult to
predict accurately, researchers proposed a combined model
to forecast. That is, according to the advantages and
Fig 4. New Requirements of Forecasting in New Situation
disadvantages of various algorithms, several algorithm
models are reasonably combined to achieve the purpose of
improving the accuracy of the prediction result. 1) More Dimensions
Under the background of spot market. The input
The combined method mainly uses both the statistics dimensions of the forecasting model are becoming more
base methods and the artificial intelligence methods to complicated. In addition to historical load information and
synthesize their advantages, in order to improve the meteorological data, the demand forecasting model should
accuracy of the final prediction. Generally speaking, there also pay attention to the electricity price and the behavior of
some ways of combination in combined mode[6]: various market participants.
 Loose combination; Meanwhile, the dimensions of the output of the
 parallel combination; forecasting model are also becoming broader. Compared
 tandem combination; with the traditional forecasting model. The new model
 network learning combination; should pay more additional attention to predict electricity
 Structural equivalence combination. price and the behaviors of all the players in the market.
These extra market information puts a big challenge on
As a result, the combined model is widely used in the
traditional forecasting model. Facing the complex situation
long-term and short-term demand forecasting.
in power maker and the behavior of each participants, more
psychology strategies should be reflected in the forecasting
IV. DEMAND FORECASTING IN THE NEW SITUATION model.
Although the electricity market in China starts late, and However, none of the current forecasting model has the
the market mechanism is not mature right now, it is same 'wisdom' as human. Fortunately, the development of
conceivable that the power market in China will be one of artificial intelligence technology makes it possible to solve
the biggest power market in the world. that question. By using artificial intelligence technology in
forecasting, the models can think like human and takes the
strategies of other participants into account. As a result,
forecasting using AI technology can grantee the accuracy of
the prediction.
2) Shorter Time
The error of demand forecast will lead to loss of
economic and energy. So, reducing the error of spot market
Fig 3. Traditional Demand Forecasting
forecasting is becoming an important issue in spot market.
That means we should focus more on short-term(very-short-
term) forecasting rather than long-term forecasting.
Considering the characteristics of the spot market,
minute-level forecasting would be more and more

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on July 29,2020 at 09:49:29 UTC from IEEE Xplore. Restrictions apply.
 important. This poses two challenges for demand Prediction Result
Output Layer
forecasting techniques. (4 nodes)
Hidden Layer
Firstly, the demand forecasting model needs to eliminate (10 nodes)
redundant data in order to simplify the input data and reduce Encoder2
… … … … Stacked
the computational complexity of the model. To achieve this (8 nodes)
Auto-Encoders
Encoder1
goal, the forecasting model should have some 'intelligence' (8 nodes) … … … … （SAE）
in order to decide which data should be taken or discarded. Input Layer
… …
The traditional data processing technology may face some (14 nodes)
Load Data Input
difficulties when dealing with this problem. However, With
Fig 5. the deep ANN stacked by Auto-Encoders for load data compression
the rapid development of the technology of big data and prediction
processing, it will make it possible to solve the problem.
After training this stacked network with test data set. We
B. Outlook of Demand Forecasting get the result as shown in Fig.6.
From the perspective of spot market, there are some
limitations when using the traditional forecasting methods. Training: R=0.99319 Validation: R=0.98312
800 800
We will then discuss several commonly used methods in Data Data
700 Fit 700 Fit
demand forecasting. Y=T Y=T
600 600

For neural network method, lack of learning or over- 500 500

fitting in the training process might lead to the failure of the 400 400

neural network. What's more, the convergence speed 300 300

becomes slow when the training data is complex, and the 200 200

solution that the model may fall into the local minima. 200 400 600 800 200 400 600 800
Target Target
For the time series method. Firstly, it requires that the
historical data is quite accurate. Secondly, time series 800
Test: R=0.98467
800
All: R=0.99

method is not sensitive to meteorological factors when it is 700

Data
Fit 700
Data
Fit
applied in short-term forecasting, which means the solution 600
Y=T
600
Y=T

of the time series method may have big error because of the 500 500
change of weather. 400 400

For regression method, it is a statistical process for 300 300

estimating the relationships among variables. This method 200 200

highly depends on the data input. Thus, when the amount of 200 400 600 800 200 400 600 800

the data becomes huge, the regression method cannot give Target Target

the solution in reasonable time. Fig 6. The performance of stacked network after training.
With the increasing of the number of smart meters in the
From Fig 6. we can get the conclusion that the accuracy
user side, on one hand, it can provide more data that
is high when using this stacked network to predict load
supports the safe operation of power system. On the other
curve. What’s more the time for training such a network just
hand, massive data also poses new challenges to load
need a few seconds.
forecasting techniques. How to meet the accuracy
requirements of the forecasting in a reasonable time is a big So, we can see that load forecasting combined with AI
problem. Fortunately, the growing artificial intelligence technology is promising in the future.
technology is expected to solve this problem in (very) short-
term forecasting model. D. Further Imagination
Load forecasting combined with AI technology will use Undoubtedly, the artificial intelligence technology will
techniques such as deep learning, machine learning and data greatly change human life. In November 2015, a
mining to select the crucial data in demand forecasting. And heavyweight study appeared on the cover of Science
give the prediction solution as accurate as possible in a magazine that artificial Intelligence finally passed the
reasonable time. The success of AlphaGo and AlphaZero Turing Test, which means that AI can think like human to
makes us believe that artificial intelligence technology will some extent. In the future, it is also expected that artificial
have a tremendous impact on demand forecasting in the near intelligence will replace some work that is doing by human
future. nowadays.
Back to the topics discussed in this article, nowadays,
C. Load Forecasting Using Deep Artificial Neural the demand forecasting is a specialized skill that should be
Network mastered by people who work in power industry. However,
Using the real load data of business users in Foshan, with the continuous development of artificial intelligence
Guangdong in 2016, we select some load data from March technology and the coming of the electricity market, all
to May as the test data set. users of power grid have the ability to do the same thing.
They can predict the price of the electricity, and then using
Using the technique of auto-encode(AE)[7], we design a that price to guide their behaviors in the electricity market,
deep artificial neural network with structure as Fig.5. such as bidding. That is to say, it means demand forecasting
will involve the linkage between multiple AI systems.

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on July 29,2020 at 09:49:29 UTC from IEEE Xplore. Restrictions apply.
 According to the principle of economics, multiple
artificial intelligence demand forecasting systems can be
regarded as many rational persons in the market. The
decisions made by these 'rational people' is to maximize
their own profits. And at the same time, the decisions they
made will also maximize the efficiency of the electricity
market. As a result, the electricity market can fulfill its
function that can allocate of power in an efficient way.

REFERENCE
[1] G. Gross and F. D. Galiana, “Short-term load forecasting,” Proc.
IEEE, vol. 75, no. 12, pp. 1558–1573, 1987.
[2] Pan Chao, Lu Baochun, Sun Liying, and Guan Weiguo, “Review of
Load Forecasting Methods in Power System,” J. Liaoning Univ.
Technol. Sci. Ed., vol. 36, no. 5, pp. 296–299, 2016.
[3] D. W. Bunn, “Forecasting loads and prices in competitive power
markets,” Proc. IEEE, vol. 88, no. 2, pp. 163–169, 2000.
[4] T. Hong and S. Fan, “Probabilistic electric load forecasting: A
tutorial review,” Int. J. Forecast., vol. 32, no. 3, pp. 914–938, 2016.
[5] A. R. Khan, A. Mahmood, A. Safdar, Z. A. Khan, and N. A. Khan,
“Load forecasting, dynamic pricing and DSM in smart grid: A
review,” Renew. Sustain. Energy Rev., vol. 54, pp. 1311–1322,
2016.
[6] Liu Jianjun, “Review of Load Forecasting Methods in Power
System,” China Sci. Technol. Inf., no. 16, pp. 52–53, 2016.
[7] P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Manzagol,
“Extracting and composing robust features with denoising
autoencoders,” in Proceedings of the 25th international conference
on Machine learning, 2008, pp. 1096–1103.

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on July 29,2020 at 09:49:29 UTC from IEEE Xplore. Restrictions apply.