AWS re:Invent 2018 - AIM401-R2 - Deep Learning Applications with Tensorflow
•Download as PPTX, PDF•
5 likes•1,016 views
The document discusses Amazon SageMaker and TensorFlow for deep learning applications. It provides an overview of SageMaker's capabilities for building, training and deploying machine learning models at scale. TensorFlow is discussed as an open source deep learning framework that is a first-class citizen on SageMaker. The document then presents a case study of how Advanced Microgrid Solutions used SageMaker and TensorFlow to develop deep learning models for forecasting electricity prices and optimizing participation in energy markets.
Recommended
More Related Content
What's hot
What's hot (20)
Similar to AWS re:Invent 2018 - AIM401-R2 - Deep Learning Applications with Tensorflow
Similar to AWS re:Invent 2018 - AIM401-R2 - Deep Learning Applications with Tensorflow (20)
More from Julien SIMON
More from Julien SIMON (20)
Recently uploaded
Recently uploaded (20)
AWS re:Invent 2018 - AIM401-R2 - Deep Learning Applications with Tensorflow
- 2. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Deep LearningApplications UsingTensorFlow A I M 4 0 1 - R 2 Julien Simon Principal Tech. Evangelist, AI/ML Amazon Web Services @julsimon Kevin Clifford Senior Product Manager Advanced Microgrid Solutions Andrew Martinez Staff Research Scientist Advanced Microgrid Solutions
- 3. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Agenda • Amazon SageMaker • TensorFlow • Case study: Advanced Microgrid Solutions • Resources
- 4. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 5. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. AmazonSageMaker Collect and prepare training data Choose and optimize your ML algorithm Set up and manage environments for training Train and tune model (trial and error) Deploy model in production Scale and manage the production environment Easily build, train, and deploy Machine Learning models
- 6. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. AmazonSageMaker Notebook instances K-Means Clustering Principal Component Analysis Neural Topic Modelling Factorization Machines Linear Learner XGBoost Latent Dirichlet Allocation Image Classification Seq2Seq, And more! ALGORITHMS Apache MXNet, Chainer TensorFlow, PyTorch Caffe2, CNTK, Torch FRAMEWORKS Set up and manage environments for training Train and tune model (trial and error) Deploy model in production Scale and manage the production environment Built-in, high- performance algorithms Build
- 7. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. AmazonSageMaker Notebook instances Built-in, high- performance algorithms One-click training Automatic Model Tuning Build Train Deploy model in production Scale and manage the production environment
- 8. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. AmazonSageMaker Fully managed hosting with auto-scaling One-click deployment Notebook instances Built-in, high- performance algorithms One-click training Automatic Model Tuning Build Train Deploy
- 9. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. SelectedAmazonSageMaker customers
- 10. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 11. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. TensorFlow • Open source software library for Machine Learning • MainAPI in Python, experimental support for other languages • Built-in support for many network architectures: FC, CNN, LSTM, etc. • Support for symbolic execution, as well as imperative execution since v1.7 (aka eager execution) • Complemented by the Keras high-levelAPI
- 12. “Of 388 projects, 80 percent using TensorFlow and other frameworks are running exclusively on AWS. 88% using only TensorFlow are running exclusively on AWS.” Nucleus Research report, December 2017 https://aws.amazon.com/tensorflow AWS is the place of choice for TensorFlow workloads
- 13. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. TensorFlowonAmazonSageMaker:afirst-classcitizen • Built-inTensorFlow containers for training and prediction • Code available on Github: https://github.com/aws/sagemaker-tensorflow-containers • Build it, run it on your own machine, customize it, etc. • Supported versions: 1.4.1, 1.5.0, 1.6.0, 1.7.0, 1.8.0, 1.9.0, 1.10.0, 1.11.0 • Advanced features • Optimized both for GPUs and CPUs (Intel MKL-DNN library) • Distributed training • Pipe mode • TensorBoard • Keras • Automatic ModelTuning
- 14. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Using Keras onAmazonSageMaker • Keras is a popular API running on top ofTF,Theano and Apache MXNet. • The tf.keras API is natively supported in Amazon SageMaker • To use Keras itself (keras.*), you need to build a custom container. • This is not difficult! • Write a Dockerfile. • Build the container. • Push it to Amazon ECR. • Use it with sagemaker.estimator.Estimator. • Full instructions and demo in this AWS Innovate talk: https://www.youtube.com/watch?v=c8Nhwr9VmfM
- 15. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Example: MNISTwithaFullyConnected network import tensorflow as tf mnist = tf.keras.datasets.mnist (x_train, y_train),(x_test, y_test) = mnist.load_data() x_train, x_test = x_train / 255.0, x_test / 255.0 model = tf.keras.models.Sequential([ tf.keras.layers.Flatten(), tf.keras.layers.Dense(512, activation=tf.nn.relu), tf.keras.layers.Dropout(0.2), tf.keras.layers.Dense(10, activation=tf.nn.softmax) ]) model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy']) model.fit(x_train, y_train, epochs=10) model.evaluate(x_test, y_test)
- 16. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. AutomaticModelTuning Finding the optimal set of hyper parameters 1. Manual Search (”I know what I’m doing”) 2. Random Search (“Spray and pray”) 3. Grid Search (“X marks the spot”) • Typically training hundreds of models • Slow and expensive 4. Hyper ParameterOptimization: use Machine Learning • Training fewer models • Gaussian Process Regression and Bayesian Optimization, https://docs.aws.amazon.com/sagemaker/latest/dg/automatic-model-tuning-how-it-works.html
- 17. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Kevin Clifford Senior Product Manager Advanced Microgrid Solutions Andrew Martinez Staff Research Scientist Advanced Microgrid Solutions
- 18. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Advanced MicrogridSolutions (AMS) Founded in 2013 in San Francisco, CA Technology-agnostic energy platform and services company that maximizes wholesale energy market revenues for both behind-the-meter and front-of-the- meter assets Control and Aggregation Economic Optimization Grid and Market Participation Situational and Business Intelligence Asset Health Management Front-of-the-MeterBehind-the-Meter AMS Transactive Energy Platform Utilities Retailers Trading Desks Customers Storage LoadControl SolarPV Storage SolarPV Hydro Wind ConventionalGeneration Markets Mission Statement: To lead a worldwide transformation to a clean energy economy by facilitating the deployment and optimization of clean energy assets
- 19. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Energy Markets 101 Demand (MW) Energy Price ($/MWh) • Electricity is traded in regional “wholesale energy markets” • Supply = Demand at all times... or grid will fail • Demand varies due to weather and behavioral factors • Market operator must procure correct amount of supply to meet demand • Suppliers must decide price and quantity to bid for every trading interval • Market Price is set at the most expensive supplier needed to meet demand
- 20. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. EnergyTechnologies 101 Less complex More complex Thermal (coal, gas, oil) Bidding at marginal cost Renewables (solar, wind) Bidding at zero marginal cost + REC value Hydroelectric Use-limited resource bidding at opportunity cost Batteries Use-limited resource bidding at opportunity cost across multiple market products
- 21. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. UseCase Optimize market participation of clean energy assets in Australia’s National Energy Market Australia’s National Energy Market (NEM) facilitates energy production for the 5 east Australian states • Serves 9M customers • $16.6B / 200TWh traded annually NEM is a “spot” market • All parties bid to consume or generate energy during upcoming 5- minute time window • 9 unique market products for energy and ancillary services
- 22. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. 4:00:30 Market data available for 4:00- 4:05 trading interval 4:03:30 Market closes for 4:05-4:10 trading interval (trading interval you want to participate in) Use Case Challenges 4:00 4:05 4:10 Left with only 180 seconds to: Forecast prices for upcoming trading intervals Determine optimal asset dispatch Construct competitive market bids Present to user for final confirmation Deliver to Market Operator …which we have to repeat every 5 minutes 0- 50s 90s 10s 20s 10s • Multiple market products • Considerations across time • Volatility (timing & magnitude) • Drift (changing market composition) • Rapidity (< 20 sec) • Frequency (5 min) • Accuracy
- 23. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. ExistingMarketForecast,whyMachineLearning? • Spot prices are forecasted by balancing generation and load bids • Power flow optimization model • Bids must be supplied through end of day, but can be updated at every market interval (5 minutes) • Market forecast accuracy is subject to the bidding behavior of market participants
- 24. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. WhyNeuralNetworks? Complex market dynamics • Seasonality and common exogenous factors, such as weather • Network outages & neighboring market conditions Increasing volatility • Changing generation portfolio with increasing penetration of (intermittent) renewable Lago, Jesus, et al. “Forecasting Spot Electricity Prices: Deep Learning Approaches and Empirical Comparison ofTraditional Algorithms.” Applied Energy, vol. 221, 2018, pp. 386–405., doi:10.1016/j.apenergy.2018.02.069. Green, Richard, and NicholasVasilakos. “Market Behaviour with Large Amounts of Intermittent Generation.” Energy Policy, vol. 38, no. 7, 2010, pp. 3211–3220., doi:10.1016/j.enpol.2009.07.038.
- 25. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. ArchitectureOverview • Data ingestion • Pre-processing • Model tuning and deployment via Amazon Sagemaker +TensorFlow + Keras • Post-processing • API wraps individual product models used for inference and scenario generation • Deployed viaAWS Chalice Data MS SQL instance ModelTuning and Deployment Forecast API Training Data, Hyperparameters, and Model Artifacts
- 26. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. AMSForecastMachineLearning Model Model design considerations 1. Learn the deviation between market forecast and historical prices 2. Multi-period forecasts of both point estimates and prediction intervals 3. Develop a framework for efficient simulation of price scenarios for stochastic optimization
- 27. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Benchmark Model • Learn the deviation between market forecast and historical prices • Input: market forecast • Output: multi-step ahead cleared market prices layers = [ tf.keras.layers.Dense( units=units, activation=activation, ), tf.keras.layers.Dropout( rate=dropout_rate, ), tf.keras.layers.Dense( units=n_forecast_intervals, ), ]
- 28. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Uncertainty Estimation • Predict both point estimates as well as uncertainty • Add an output dimension to represent quantiles layers = [ tf.keras.layers.Dense( units=units, activation=activation, ), tf.keras.layers.Dropout( rate=dropout_rate, ), tf.keras.layers.Dense( units=n_forecast_intervals * n_quantiles, ), tf.keras.layers.Reshape( target_shape=(n_forecast_intervals, n_quantiles), ), ] layers = [ tf.keras.layers.Dense( units=units, activation=activation, ), tf.keras.layers.Dropout( rate=dropout_rate, ), tf.keras.layers.Dense( units=n_forecast_intervals, ), ]
- 29. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Uncertainty Estimation • Quantile Regression • Asymmetric (pinball) loss function conditional to quantile, 𝜏 ∈ 0, 1 Example: Under-prediction, −10 = 9 Over-prediction, 10 = 1 tf.losses.compute_weighted_loss( losses=tf.maximum( -quantile * error, (1– quantile) * error ), weights=weights, reduction=Reduction.MEAN, )
- 30. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Example Forecast • Single inference • 300 forecast intervals • 5 quantiles: 10, 30, 50, 70, 90%
- 31. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. StochasticScenarioGeneration • Develop a framework for efficient simulation of price scenarios for stochastic optimization • Use test set to derive temporal covariance • Now sample from a known distribution to generate realistic price scenarios! np.random.multivariate_normal( mean=np.zeros(n_forecast_intervals), cov=covariance, size=n_scenarios )
- 32. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Benchmark Model Results Meets requirements • Improvement on market forecasts • Single model, capable of quick quantile estimation and scenario generation Limitations • Densely connected model prone to overfitting when additional features are included Next steps • Develop a more robust model, less dependent on regularization methods • Use intuition for feature & target dependencies to reduce model connectivity
- 33. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. ConvolutionalNetwork Dilated convolutional neural network • Example number of layers = 3 kernel size = 3 dilation rate = 3 • Receptive field grows exponentially • Captures both short and long-term dependencies • Typically stacked with residual connections
- 34. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Hyperparameter and ModelTuning Data parameters • Missing value imputation method • Sample weight exponential decay rate Model architecture • Filter size • Kernel size • Dropout rate • Dilation rate • Number of layers (and stacks) • Skip connections • Causal vs acausal convolution Solver parameters • Learning rate tuner = HyperparameterTuner( estimator=self.estimator, objective_metric_name='validation-loss', hyperparameter_ranges=hyperparameter_ranges, metric_definitions=[{ 'Name': 'validation-loss', 'Regex': ', loss = ([+-]?d+.d+)', }], strategy='Bayesian', objective_type='Minimize', max_jobs=max_jobs, max_parallel_jobs=max_parallel_jobs, tags=tags, base_tuning_job_name=self.job_name, )
- 35. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. HyperparameterTuning Tuning example • 36 jobs (3 in parallel) • 2 hours per training job • 72 training hours • ml.p2.xlarge,Tesla K80, $1.26/hr • 20% difference in loss among jobs
- 36. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Accuracy • We follow common model validation practices, splitting the data into three chronologically separated groups train – model training validate – hyperparameter tuning test – final comparison metric, empirical error used to derive covariance matrix, and estimation of future performance • Case Study: South Australia energy prices, 24-hour point estimates 68% reduction of mean absolute error (against market forecast) 24% reduction of median absolute error
- 37. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Takeaways Without prior experience using Deep Learning tools • Deployed benchmarkTensorFlow model on AWS in weeks • Learned behavioral market patterns improves upon market-generated forecast • Extended the model to state-of-the-art temporal dilated convolutional network • Single forecast API provides quantile forecasts and realistic product- temporal-correlated price scenarios for all queried products
- 38. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Recommendations • Start simple, extending the examples provided by AWS https://github.com/awslabs/amazon-sagemaker-examples • Keras allows for quick prototyping • Parameterize the “art” of Deep Learning architecture and let tuning discover best design
- 39. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Presenters Kevin Clifford Andrew Martinez Corresponding author Corey Noone (coreyn@advmicrogrid.com) Shameless plug We’re hiring data scientists and software developers Advanced Microgrid Solutions 986 Mission St, 4th Floor San Francisco, CA 94103 https://advmicrogrid.com
- 40. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 41. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Resources https://ml.aws https://tensorflow.org/ https://keras.io/ https://aws.amazon.com/sagemaker https://github.com/awslabs/amazon-sagemaker-examples https://github.com/aws/sagemaker-python-sdk https://medium.com/@julsimon
- 42. © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
- 43. Thank you! © 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved. Julien Simon PrincipalTech. Evangelist, AI/ML AmazonWeb Services @julsimon Kevin Clifford Senior Product Manager Advanced Microgrid Solutions Andrew Martinez Staff Research Scientist Advanced Microgrid Solutions
Editor's Notes
- Amazon SageMaker removes the complexity that holds back developer success with each of these steps. Amazon SageMaker includes modules that can be used together or independently to build, train, and deploy your machine learning models.
- SageMaker makes it easy to build ML models and get them ready for training by providing everything you need to quickly connect to your training data, and to select and optimize the best algorithm and framework for your application. Amazon SageMaker includes hosted Jupyter notebooks that make it is easy to explore and visualize your training data stored in Amazon S3. You can connect directly to data in S3, or use AWS Glue to move data from Amazon RDS, Amazon DynamoDB, and Amazon Redshift into S3 for analysis in your notebook. To help you select your algorithm, Amazon SageMaker includes the 10 most common machine learning algorithms which have been pre-installed and optimized to deliver up to 10 times the performance you’ll find running these algorithms anywhere else. Amazon SageMaker also comes pre-configured to run TensorFlow and Apache MXNet, two of the most popular open source frameworks, or you have the option of using your own framework.
- You can begin training your model with a single click in the Amazon SageMaker console. The service manages all of the underlying infrastructure for you and can easily scale to train models at petabyte scale. To make the training process even faster and easier, Amazon SageMaker can automatically tune your model to achieve the highest possible accuracy.
- Once your model is trained and tuned, SageMaker makes it easy to deploy in production so you can start generating predictions on new data (a process called inference). Amazon SageMaker deploys your model on an auto-scaling cluster of Amazon EC2 instances that are spread across multiple availability zones to deliver both high performance and high availability. It also includes built-in A/B testing capabilities to help you test your model and experiment with different versions to achieve the best results. For maximum versatility, we designed Amazon SageMaker in three modules – Build, Train, and Deploy – that can be used together or independently as part of any existing ML workflow you might already have in place.
- https://www.businesswire.com/news/home/20180404006122/en/Tens-Thousands-Customers-Flocking-AWS-Machine-Learning Edmunds.com is a car-shopping website that offers detailed, constantly updated information about vehicles to 20 million monthly visitors. “We have a strategic initiative to put machine learning into the hands of all our engineers,” said Stephen Felisan, Chief Information Officer at Edmunds.com. “Amazon SageMaker is key to helping us achieve this goal, making it easier for engineers to build, train, and deploy machine learning models and algorithms at scale. We are excited to see how we can use Amazon SageMaker to innovate new solutions across the organization for our customers.” The Move, Inc. network, which includes Realtor.com, Doorsteps, and Moving.com, provides real estate information, tools, and professional expertise across a family of websites and mobile experiences for consumers and real estate professionals. “We believe that Amazon SageMaker is a transformative addition to the realtor.com toolset as we support consumers along their homeownership journey," said Vineet Singh, Chief Data Officer and Senior Vice President at Move, Inc. "Machine learning workflows that have historically taken a long time, like training and optimizing models, can be done with greater efficiency and by a broader set of developers, empowering our data scientists and analysts to focus on creating the richest experience for our users." Dow Jones is a publishing and financial information firm that publishes the world's most trusted business news and financial information in a variety of media. It delivers breaking news, exclusive insights, expert commentary and personal finance strategies. “As Dow Jones continues to focus on integrating machine learning into our products and services, AWS has been a great resource,” said Ramin Beheshti, Group Chief Product and Technology Officer. “Leading up to our recent Machine Learning Hackathon, the AWS team provided training to participants on Amazon SageMaker and Amazon Rekognition, and offered day-of support to all the teams. The result was that our teams developed some great ideas for how we can apply machine learning, many of which we we’ll continue to develop on AWS. The event was a huge success, and an example of what a great relationship can look like.” Every day Grammarly’s algorithms help millions of people communicate more effectively by offering writing assistance on multiple platforms across devices. Through a combination of natural language processing and advanced machine learning technologies, Grammarly is tackling critical communication and business challenges. “Amazon SageMaker makes it possible for us to develop our TensorFlow models in a distributed training environment,” said Stanislav Levental, Technical Lead at Grammarly. “Our workflows also integrate with Amazon EMR for pre-processing, so we can get our data from Amazon Simple Storage Service (Amazon S3), filtered with Amazon EMR and Spark from a Jupyter notebook, and then train in Amazon SageMaker with the same notebook. Amazon SageMaker is also flexible for our different production requirements. We can run inferences on Amazon SageMaker itself, or if we need just the model, we download it from Amazon S3 and run inferences of our mobile device implementations for iOS and Android customers.” Cookpad is Japan’s largest recipe sharing service, with about 60 million monthly users in Japan and about 90 million monthly users globally. “With the increasing demand for easier use of Cookpad’s recipe service, our data scientists will be building more machine learning models in order to optimize the user experience,” said Mr. Yoichiro Someya, Research Engineer at Cookpad. “Attempting to minimize the number of training job iterations for best performance, we recognized a significant challenge in the deployment of machine learning inference endpoints, which was slowing down our development processes. To automate the machine learning model deployment such that data scientists could deploy models by themselves, we used Amazon SageMaker inference APIs and proved that Amazon SageMaker would eliminate the need for application engineers to deploy machine learning models. We anticipate automating this process with Amazon SageMaker in production.”
- Image source: https://commons.wikimedia.org/wiki/File:MnistExamples.png
- Image source: https://en.wikipedia.org/wiki/File:Hypercube.svg
- Before diving in to the use case, I want to establish some fundamental concepts around the “industry-specific” terms used.
- Electricity is traded in regional “wholesale energy markets”, for example there is a market that covers California. One of the primary purposes of this financial market to ensure that supply exactly equals demand at all times, as failing to do so would disrupt the physical power systems which could result in grid failure and wide-spread power outages. This element of “at all times” is complicated by large variations in demand through time generally dictated by changes behavior and environmental factors like weather. So the market operator role is to procure the right amount of energy supply to match this changing target. For suppliers participating in this market, they bids into the market the quantity that they are willing to generate for at different price points (more on how they select their price point in a moment). The outcome of this matching process is the market clearing price, which is for the most expensive energy source procured to meet demand, and is the price paid to to all suppliers. All energy is bought and sold in regional “wholesale energy markets”. As demand and supply varies dramatically through time, so does the corresponding price of energy.
- There are several types of technologies that generate and sell energy that end-users ultimately buy and consume. These include thermal (gas-fired) generators, renewable resources like wind and solar, hydroelectric dams, and the new kid on the block – battery energy storage. Each of these differ in the manner they generate power in but also have very different costs to generate (comparable to cost of goods sold). Some technology-type’s costs are rather straight forward, thermal generators for example are subject to the price of gas burned to generate electricity. Others are much more complex. Batteries costs to generate come in the form of charging and therefore must make price arbitrage decisions of charging cost to discharge revenue. Simplistically this means discharging during high prices to maximize revenue and charge during lower prices to minimize cost. Accurate understanding of future prices is critical to correct decision making for profitable energy market participation. Image Credits Coal Plant: https://commons.wikimedia.org/wiki/File:Sherco_Generating_Station_-_Xcel_Energy_Sherburne_County_Coal-Fired_Power_Plant_-_Sunset_(24077210421).jpg Solar Farm: https://commons.wikimedia.org/wiki/File:Taean_Solar_Farm_at_7pm_-_panoramio.jpg Wind Farm: https://commons.wikimedia.org/wiki/File:Sunset_at_Royd_Moor_Wind_Farm.jpg Dam: https://commons.wikimedia.org/wiki/File:Diablo_Dam_(from_WA_SR_20).jpg
- Our use case is to optimize the market participation or manner of market bidding for energy storage assets in Australia’s National Energy Market. This market serves 9M customers spread around the five eastern AUS states. This is one of the largest energy markets in the world both in terms of volume and value, where 200 TWh and $16.6B are traded annually. The Australia market is designed as a “spot” market where all supply and demand bid to generate and consume energy during the upcoming 5-minute time window. Additionally there are 9 different avenue in which you can bid. In practice all parties submit a bid (quantity & price) to the market. The market operator considers all bids, awards the suppliers with the best bids the ability to generate during this time window. Wrapping in batteries, value of discharging at a later point in time has to be weigh against discharging now. So while the upcoming 5-minutes is the focus of the market, proper bidding of the battery requires an understanding of future time intervals as well. Image Credit: https://www.electranet.com.au/what-we-do/network/national-electricity-market-and-rules/
- Provide motivation for developing our solution before getting into the how we are implementing the solution AEMO is unique in that it publishes it’s own market forecast
- Didn’t know if any or all of these features are important, which makes it well suited for neural networks, as the model will decide what is important The complexity and volatility of the market is also increases so want a framework that would work well into the future There were also a few papers that showed promising results
- Before getting in to the model, want to take a step back to look at the components that we ended up using and briefly describe the final architecture.
- Walk through the development from simple model to the more complicated, highly parameterized model.
- Previous model will give you the point estimate, but we also want to model uncertainty. This can be done by adding a dimension to the output that represents the quantiles.