Returning to the essence, AIGC’s breakthrough in singularity is a combination of three factors:

1. GPT is a replica of human neurons

GPT AI represented by NLP is a computer neural network algorithm, whose essence is to simulate neural networks in the human cerebral cortex.

The processing and intelligent imagination of language, music, images, and even taste information are all functions accumulated by the human

brain as a “protein computer” during long-term evolution.

Therefore, GPT is naturally the most suitable imitation for processing similar information, that is, unstructured language, music, and images.

The mechanism of its processing is not the understanding of meaning, but rather a process of refining, identifying, and associating. This is a very

paradoxical thing.

Early speech semantic recognition algorithms essentially established a grammar model and a speech database, then mapped the speech to the vocabulary,

then placed the vocabulary into the grammar database to understand the meaning of the vocabulary, and finally obtained recognition results.

The recognition efficiency of this “logical mechanism” based syntax recognition has been hovering around 70%, such as the ViaVoice recognition

algorithm introduced by IBM in the 1990s.

AIGC is not about playing like this. Its essence is not to care about grammar, but rather to establish a neural network algorithm that allows the

computer to count the probabilistic connections between different words, which are neural connections, not semantic connections.

Much like learning our mother tongue when we were young, we naturally learned it, rather than learning “subject, predicate, object, verb, complement,”

and then understanding a paragraph.

This is the thinking model of AI, which is recognition, not understanding.

This is also the subversive significance of AI for all classical mechanism models – computers do not need to understand this matter at the logical level,

but rather identify and recognize the correlation between internal information, and then know it.

For example, the power flow state and prediction of power grids are based on classical power network simulation, where a mathematical model of the

mechanism is established and then converged using a matrix algorithm. In the future, it may not be necessary. AI will directly identify and predict a

certain modal pattern based on the status of each node.

The more nodes there are, the less popular the classical matrix algorithm is, because the complexity of the algorithm increases with the number of

nodes and the geometric progression increases. However, AI prefers to have very large scale node concurrency, because AI is good at identifying and

predicting the most likely network modes.

Whether it’s the next prediction of Go (AlphaGO can predict the next dozens of steps, with countless possibilities for each step) or the modal prediction

of complex weather systems, AI’s accuracy is much higher than that of mechanical models.

The reason why the power grid currently does not require AI is that the number of nodes in 220 kV and above power networks managed by provincial

dispatching is not large, and many conditions are set to linearize and sparse the matrix, greatly reducing the computational complexity of the

mechanism model.

However, at the distribution network power flow stage, facing tens of thousands or hundreds of thousands of power nodes, load nodes, and traditional

matrix algorithms in a large distribution network is powerless.

I believe that pattern recognition of AI at the distribution network level will become possible in the future.

2. The accumulation, training, and generation of unstructured information

The second reason why AIGC has made a breakthrough is the accumulation of information. From the A/D conversion of speech (microphone+PCM

sampling) to the A/D conversion of images (CMOS+color space mapping), humans have accumulated holographic data in the visual and auditory

fields in extremely low-cost ways over the past few decades.

In particular, the large-scale popularization of cameras and smartphones, the accumulation of unstructured data in the audiovisual field for humans

at almost zero cost, and the explosive accumulation of text information on the Internet are the key to AIGC training – training data sets are inexpensive.

The figure above shows the growth trend of global data, which clearly presents an exponential trend.

This non-linear growth of data accumulation is the foundation for the non-linear growth of AIGC’s capabilities.

BUT, most of these data are unstructured audio-visual data, which is accumulated at zero cost.

In the field of electric power, this cannot be achieved. Firstly, most of the electric power industry is structured and semi structured data, such as

voltage and current, which are point data sets of time series and semi structured.

Structural data sets need to be understood by computers and require “alignment”, such as device alignment – the voltage, current, and power data

of a switch need to be aligned to this node.

More troublesome is time alignment, which requires aligning voltage, current, and active and reactive power based on the time scale, so that

subsequent identification can be performed. There are also forward and reverse directions, which are spatial alignment in four quadrants.

Unlike text data, which does not require alignment, a paragraph is simply thrown to the computer, which identifies possible information associations

on its own.

In order to align this issue, such as the equipment alignment of business distribution data, alignment is constantly needed, because the medium and

low voltage distribution network is adding, deleting, and modifying equipment and lines every day, and grid companies spend huge labor costs.

Like “data annotation,” computers cannot do this.

Secondly, the cost of data acquisition in the power sector is high, and sensors are required instead of having a mobile phone to speak and take photos. ”

Every time the voltage decreases by one level (or the power distribution relationship decreases by one level), the required sensor investment increases

by at least one order of magnitude. To achieve load side (capillary end) sensing, it is even more a massive digital investment.”.

If it is necessary to identify the transient mode of the power grid, high-precision high-frequency sampling is required, and the cost is even higher.

Due to the extremely high marginal cost of data acquisition and data alignment, the power grid is currently unable to accumulate sufficient non-linear

growth of data information to train an algorithm to reach the AI singularity.

Not to mention the openness of data, it is impossible for a power AI startup to obtain these data.

Therefore, before AI, it is necessary to solve the problem of data sets, otherwise general AI code cannot be trained to produce a good AI.

3. Breakthrough in computational power

In addition to algorithms and data, the singularity breakthrough of AIGC is also a breakthrough in computational power. Traditional CPUs are not

suitable for large-scale concurrent neuronal computing. It is precisely the application of GPUs in 3D games and movies that makes large-scale parallel

floating-point+streaming computing possible. Moore’s Law further reduces the computational cost per unit of computational power.

Power grid AI, an inevitable trend in the future

With the integration of a large number of distributed photovoltaic and distributed energy storage systems, as well as the application requirements of

load side virtual power plants, it is objectively necessary to conduct source and load forecasting for public distribution network systems and user

distribution (micro) grid systems, as well as real-time power flow optimization for distribution (micro) grid systems.

The computational complexity of the distribution network side is actually higher than that of the transmission network scheduling. Even for a commercial

complex, there may be tens of thousands of load devices and hundreds of switches, and the demand for AI based micro grid/distribution network operation

control will arise.

With the low cost of sensors and the widespread use of power electronic devices such as solid-state transformers, solid-state switches, and inverters (converters),

the integration of sensing, computing, and control at the edge of the power grid has also become an innovative trend.

Therefore, the AIGC of the power grid is the future. However, what is needed today is not to immediately take out an AI algorithm to make money,

Instead, first address the data infrastructure construction issues required by AI

In the upsurge of AIGC, there needs to be sufficient calm thinking about the application level and future of power AI.

At present, the significance of power AI is not significant: for example, a photovoltaic algorithm with a prediction accuracy of 90% is placed in the spot market

with a trading deviation threshold of 5%, and the algorithm deviation will wipe out all trading profits.

The data is water, and the computational power of the algorithm is a channel. As it happens, it will be.